TL;DR: You automate commercial real estate acquisitions stage by stage, not all at once: inbound deal screening against your buy box, OM and rent roll/T12 intake, comps assembly, first-pass underwriting, the IC memo first draft, and pipeline tracking. AI is now good at the reading, extracting and reconciling work in each stage, and per Dealpath's 2025 survey, 49% of institutional investors already use it for offering memorandum extraction. The firms that see real time savings are the ones that connect those stages to their own deal history and portfolio data, measure one countable unit (deals screened per analyst per week), and keep the pricing and go/no-go calls with people.

It is Monday morning and the acquisitions inbox has 41 new emails from brokers. Eleven have OMs attached. Three have a rent roll and a T12 in a data room link. One is a 9-acre parcel a landowner's nephew sent in a text message photo. Your associate will spend the next two days retyping rent rolls into the model and Googling submarket comps, and by Thursday you will have looked hard at maybe four of those deals. The other 37 get a skim, a gut call, or nothing.

That is the problem acquisitions automation solves. Not "AI that picks deals," but a team that sees every deal that hits the inbox, gets a consistent first read on each one within hours, and spends analyst time on the handful that deserve a full model. This guide is for Heads of Acquisitions, principals, and CEOs at CRE owner/operators and investment firms who want to know what actually works today, stage by stage, what it costs in effort, and where the human stays in charge.

If you want the conceptual background on agents and knowledge graphs in CRE, we covered that in AI agents and knowledge graphs for commercial real estate. This post is the operator's playbook: the specific acquisition stages, the documents involved, what to automate first, and how to prove it paid off.

Key takeaways

• Extraction is already mainstream. Per Dealpath's 2025 survey of 100 institutional investors, 49% use AI for OM extraction or review, 43% for AI-recommended comps, and 40% for rent roll and lease extraction.

• Pilots are common, results are rare. JLL's 2025 Global Real Estate Technology Survey found 88% of investors, owners and landlords have started AI pilots, yet only 5% of companies report achieving all of their AI program goals.

• The data layer is the bottleneck. Every firm in the Dealpath survey (100%) said its data is stored across multiple platforms, and 98% called improving data infrastructure a top priority.

• Name the countable unit first. Deals screened per analyst per week, hours from OM receipt to first-pass model, and days from LOI to IC are the metrics that make ROI provable.

• Automate in pipeline order, starting with screening. The top of the funnel has the highest volume and the lowest cost of a wrong answer, so it pays back fastest.

• Land sourcing is automatable too. County appraiser data, deed records, zoning, and federal flood and wetlands layers can produce a ranked parcel list before anyone drives a site.

• Humans keep the judgment calls. Pricing, the go/no-go, seller relationships and the final IC recommendation stay with people. Agents take the reading, reconciling and first drafts.

How do you automate commercial real estate acquisitions?

You automate commercial real estate acquisitions by breaking the process into its repeatable stages, automating the reading and data-entry work in each one, and connecting the stages so information captured once flows through to the model, the memo and the pipeline. The stages, in order, are sourcing and screening, document intake, comps, first-pass underwriting, IC memo drafting, and pipeline tracking.

The mistake most firms make is buying a point tool for one stage, usually OM extraction, and calling it done. The extraction works. The analyst still copies the output into the ARGUS or Excel model by hand, still rebuilds the comp set from scratch, and still writes the IC memo in Word from memory. The time saved in one stage leaks out at every handoff.

Here is the full acquisitions workflow and what automation looks like in each piece:

1. Sourcing and screening. Read every inbound broker email and OM, extract the basics (asset type, units or square feet, location, asking price, in-place occupancy), score against your buy box, and route.

2. Document intake. Turn the OM, rent roll, T12, and lease abstracts into structured data that matches your model's line items.

3. Comps. Pull sale and rent comps from your licensed data (CoStar, Real Capital Analytics, Yardi Matrix) and, more valuable, from your own deal history.

4. First-pass underwriting. Populate your standard model with in-place numbers and house assumptions, then flag where the broker's pro forma departs from the T12.

5. IC memo drafting. Produce a first draft in your firm's format, with every number linked back to its source page.

6. Pipeline tracking. Update deal status, dates, and next actions automatically, so the Monday pipeline meeting runs off live data instead of a spreadsheet someone updated Friday.

Our position, and it is an arguable one: if you can only automate one stage this year, automate screening, not underwriting. Underwriting automation saves an analyst hours on deals you were already going to look at. Screening automation changes which deals you look at. The second one moves returns.

How do you automate CRE deal screening against your buy box?

Automated CRE deal screening means an agent reads each inbound opportunity, extracts the handful of facts your buy box cares about, and sorts the deal into pursue, watch, or pass, with a one-paragraph reason. The analyst reviews the sort, not the raw inbox.

Write the buy box down as rules a machine can apply

Most buy boxes live in a principal's head and a slide in the fund deck. To automate screening, it has to become explicit. A typical multifamily value-add buy box might read:

• Garden or mid-rise, 1990 or newer, 150 to 400 units

• Target MSAs list, with excluded submarkets named

• In-place rents at least 10% below renovated comps

• Check size between a stated floor and ceiling

• No ground lease, no rent-controlled units above a set share

• Assumable debt welcome, flagged as a positive

Industrial, retail, and office buy boxes look different (clear height, WALT, anchor tenant credit, parking ratio), but the principle holds. If two people on your team would screen the same OM differently, fix that before you automate it.

Score against your own history, not just the rules

Rules catch the obvious passes. The better signal comes from your own records: have you seen this asset before, at what price, and why did you pass? Did you lose a similar deal in the same submarket last year by $2M? Is the broker someone whose OMs historically overstate occupancy? That context sits in old emails, CRM notes, and the shared drive, and it is exactly what a new analyst does not know. An agent that can see it screens like a ten-year veteran instead of a first-year.

What good output looks like

Each deal gets a short card: the extracted facts, the buy-box result, a pursue/watch/pass recommendation, the one or two reasons, and any history ("We bid this in 2023 at a 5.4% cap and lost to a 1031 buyer"). The analyst can override with one click, and the override is logged, because overrides are how the screening logic gets better.

Can AI read an offering memorandum, rent roll and T12?

Yes. Reading OMs, rent rolls and T12s is the most mature part of AI for commercial real estate acquisitions, and it is where most firms start. The open question is no longer whether the extraction works but whether the extracted data lands somewhere useful.

The labor being replaced is real. V7 Labs, a document AI vendor, estimates that a 200-unit multifamily rent roll takes three to four hours to abstract by hand. Multiply that across every deal that makes it past screening and you have a full-time analyst whose job is retyping.

What each document gives you

• Offering memorandum. The broker's story: property description, unit mix, amenity list, the pro forma, the sales comps the broker chose, and the market narrative. Useful, and written to sell.

• Rent roll. The current state of every unit or suite: tenant, lease start and end, in-place rent, market rent, concessions, deposits, status. For commercial assets, also options, escalations, and recovery structures.

• T12 (trailing twelve months operating statement). What the property actually earned and spent over the last year, month by month. The rent roll tells you what the property should collect today. The T12 tells you what it did collect.

Rent roll vs T12: the reconciliation is the work

People search "T12 vs rent roll" as if they compete. They are two halves of a check. Annualize the rent roll's in-place rents, compare to the T12's gross potential rent and effective rental income, and the gap tells you about vacancy, concessions, bad debt, and whether the seller's numbers hang together. A first-pass agent should run that reconciliation automatically and flag anything outside tolerance, for example "T12 shows $4.1M in rental income; annualized rent roll implies $4.6M; 11% gap suggests concessions or collections issues not mentioned in the OM."

Map to your chart of accounts, not the seller's

Every seller's T12 uses a different chart of accounts. Some export from Yardi, some from MRI, some from AppFolio, and some from an Excel file the property manager maintains. "R&M," "Repairs - General," and "Turn Costs" might all need to roll into your model's repairs and maintenance line. This mapping is where generic extraction tools fall short and where a system trained on your own model's line items earns its keep. We wrote more about the cross-system mapping problem in reconciling Yardi and ARGUS data.

How do you automate CRE comps?

You automate comps by having an agent pull candidate sale and rent comps from your licensed data sources and your own closed and lost deals, filter them by the rules your team uses (distance, vintage, size, date), and present a draft comp set with the reasoning for each inclusion. A human then accepts, removes, or adds.

Per Dealpath's 2025 survey, 43% of institutional investors already use AI-recommended comps, which tells you this is moving from experiment to standard practice. The difference between a useful comp tool and a noisy one comes down to two things.

• Your own deal history counts. The best comps are often deals you underwrote yourself, because you know the real T12, not the broker's version. If you bid on a property two miles away last spring and saw its rent roll, that comp beats anything in a third-party database. Most firms never use this data because it is buried in old Excel models on a shared drive.

• Comp selection needs to show its work. An IC member will ask why a comp is in the set. The answer has to be on the page: "Same vintage, 1.4 miles, sold 7 months ago, similar unit mix, excluded the 2024 trade because it was a portfolio sale with allocated pricing."

Can AI underwrite a commercial real estate deal?

AI can produce a credible first-pass underwrite: in-place NOI from the T12, a rent roll-based revenue build, your house assumptions for growth, capex, and exit cap, and a returns summary. It should not produce the final underwrite. The final underwrite is a set of judgments about the future that belongs to the people accountable for the outcome.

Per Dealpath's 2025 survey, 40% of institutional investors use AI for underwriting or modeling support, and 61% expect AI to deliver faster deal evaluation and closings. The way to get there without losing control is to separate two kinds of work that usually happen in the same Excel session.

The mechanical layer (automate it)

• Populating the model's inputs from the extracted rent roll and T12

• Normalizing the T12: removing one-time items, flagging below-the-line expenses the seller pushed out of NOI

• Re-assessing property taxes at the purchase price, using the county's millage rate

• Applying house assumptions (management fee, replacement reserves, insurance per unit) consistently

• Running standard sensitivities on exit cap, rent growth, and interest rate

• Flagging every place the broker's pro forma departs from in-place performance

The judgment layer (keep it human)

• What rent premium a renovation will actually earn in this submarket

• Where the exit cap should sit given today's debt market

• How much to believe the seller's story about last year's bad debt

• What the business plan is, and whether your team can execute it

The output of a good first-pass underwrite is not a number. It is a model and a list of questions. "The pro forma assumes 96% occupancy; the trailing six months averaged 89%; the rent roll shows 14 units on month-to-month. Confirm with the property manager." That list is what makes the first call with the broker productive. Our underwriting and diligence agent is built around that pattern: draft the model, surface the discrepancies, and leave the assumptions to the analyst.

Can ChatGPT or Claude underwrite a deal?

A general chatbot can read an OM you paste into it and summarize it well. It cannot see your model, your house assumptions, your past deals, or your portfolio's actual operating numbers, so its underwrite reflects the broker's view of the world plus generic defaults. That is fine for a quick read on a single deal. It is not a system, and it leaves no audit trail an IC can review.

Can AI write an investment committee memo?

Yes, AI can write a solid first draft of an IC memo, and per Dealpath's 2025 survey, 56% of institutional investors already use AI to automate investment memos. The draft is valuable when it follows your firm's template, pulls every number from the live model, and cites the source document and page for each claim.

A strong first draft covers the sections your IC expects: the investment thesis, property and market summary, the business plan, sources and uses, the returns summary with sensitivities, comp tables written as lists, key risks and mitigants, and diligence items still open. The analyst then rewrites the thesis and the risks in their own words, because those are the parts the committee is actually voting on.

Two rules make AI-drafted IC memos trustworthy:

1. Every number traces back. If the memo says "in-place NOI of $3.2M," a reader can click through to the T12 line items and the adjustments that produced it. No orphan numbers.

2. The draft says what it does not know. Open diligence items are listed, not glossed over. A memo that sounds confident about a property condition report nobody has read yet is worse than no memo.

How do you automate land and parcel sourcing?

You automate land sourcing by pulling public parcel data from the county property appraiser, joining it to deed records, zoning, and federal environmental layers, and ranking parcels that fit your development criteria before anyone calls an owner. For developers and build-to-rent, industrial, and self-storage buyers, this is often the highest-return automation of all, because off-market land is found, not listed.

The public data stack

• County property appraiser. Parcel ID, owner of record, mailing address, acreage, land use code, assessed and just values, and building improvements. Most counties publish this as a downloadable file or a GIS service.

• Deed and recorder records. Last sale date and price, deed type, and whether the owner is an LLC, a trust, an estate, or an individual. Long hold periods, out-of-state owners, and estate transfers are classic motivation signals.

• Zoning and future land use. Current zoning district and the future land use designation in the county's long-range plan. A parcel zoned agricultural with a future land use of mixed-use is a different opportunity than one locked in rural residential.

• Floodplain. FEMA's National Flood Hazard Layer, which per FEMA covers over 90% of the U.S. population, shows effective flood zones at the parcel level. A site that is 60% in Zone AE needs a different yield calculation.

• Wetlands. The U.S. Fish and Wildlife Service's National Wetlands Inventory maps wetland and deepwater habitats. It is a screening layer, not a jurisdictional delineation, so a hit means "send an environmental consultant," not "pass."

• Utilities and access. Distance to water and sewer, road frontage, and curb cuts, often from county or utility GIS.

What the agent does with it

An agent applies your site criteria (minimum net developable acres, zoning or rezoning path, maximum flood and wetland share, distance to an interchange or a target employer) and calculates net usable acreage by subtracting the flood and wetland overlap. It then produces a ranked list with an owner profile for each parcel. The output is a short list of owners worth a letter or a call, with the reasons attached. The human part, which no software replaces, is the conversation with a landowner whose family has held the property for 40 years.

One caution: county data is messy. Owner names are spelled differently across the appraiser file and the deed index, parcels get split and renumbered, and LLCs hide the real decision maker. Resolving those records into a single owner and a single parcel history is the unglamorous work that determines whether the list is any good. Our deal origination agent treats that resolution as step one.

How do you track an acquisition pipeline without spreadsheets?

You track the acquisition pipeline automatically by letting the system update deal status from the events that already happen: an OM arrives, a model is built, an LOI is sent, a PSA is signed, diligence items close. The pipeline view then reflects reality without anyone maintaining it by hand.

Most acquisitions teams run their pipeline in Excel, in Dealpath, in Salesforce, or in some mix of all three, and all of them share one weakness: somebody has to update them. The Friday update gets skipped during a busy week, and the Monday meeting runs on stale data.

An automated pipeline should answer these questions on demand:

• How many deals did we screen this week, and how many advanced?

• Which deals have a bid date in the next ten days with no model started?

• What is our hit rate by broker, by market, and by asset type over the last 12 months?

• Why did we lose the last ten deals we bid on (price, terms, timing, certainty)?

• Which diligence items are blocking deals under PSA?

That last-ten-losses question is the one most firms cannot answer from their systems today, and it is the one that would most change their bidding.

What should you not automate in CRE acquisitions?

Do not automate the decisions that carry accountability: price, the go/no-go, the business plan, and the relationships. Automate the reading, reconciling and drafting that feed those decisions.

• Pricing and the bid. An agent can show you the sensitivity table. Where you land on it reflects your cost of capital, your conviction, and what you know about the seller. That is a principal's call.

• Broker and seller relationships. Brokers send their best deals to buyers who close and who call back. An automated pass email that arrives four minutes after the OM tells a broker you did not read it.

• Site visits and property condition. No model sees the foundation crack, the smell in the hallway, or the neighbor's new construction.

• Final IC recommendation. The memo can be drafted by an agent. The recommendation is signed by a person.

• Anything legal. PSA negotiation, title objections, and environmental findings need counsel and consultants, with the agent's role limited to tracking and summarizing.

A useful test: if a wrong answer from the agent would be caught by the next human step at low cost, automate it. If a wrong answer would flow straight into a bid or a signed document, keep a person in the loop with the agent as a checker.

What data layer does acquisitions automation need underneath?

Acquisitions automation needs a connected view of your deal history, your portfolio's actual operating data, your models, and your market data, with the same property, tenant, and owner recognized as the same thing across every source. Without that layer, each agent works from one document at a time and cannot apply what your firm already knows.

This is the part most CRE AI projects skip, and it is why they stall. Per Gartner's February 2025 analysis, organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026. In CRE, the problem is specific and familiar:

• The same property is "Oakmont Apartments" in the CRM, "Oakmont Apts LLC" in the deed record, and a parcel number in the appraiser file.

• Your owned assets' actual expenses sit in Yardi or MRI, your acquisition models sit in Excel or ARGUS on a shared drive, and your comps sit in CoStar. None share an ID.

• "NOI" in the asset management report and "NOI" in the acquisition model are calculated differently, because one includes replacement reserves and one does not.

The fix is a layer that connects those systems, resolves entities (this property, this owner, this tenant are the same across sources), and keeps the relationships between them: this asset, bought in this deal, underwritten with these assumptions, now performing at these actuals. That is what we mean by a company brain, built on a knowledge graph. For an acquisitions team, the payoff is concrete. When the agent underwrites a new deal, it can check the insurance assumption against what you actually paid per unit on the six similar assets you own, instead of a default from a template.

You do not need to connect everything on day one. Start with the deal history (past models and IC memos) and the portfolio actuals from your property management system. Those two sources make screening and underwriting noticeably smarter.

How do you measure ROI on acquisitions automation?

You measure ROI by naming one countable unit before you start, recording its baseline, and tracking it weekly. For acquisitions, the best units are deals screened per analyst per week, hours from OM receipt to first-pass model, and days from LOI to IC approval.

This matters because the industry is full of pilots that never show results. MIT's Project NANDA reported in its July 2025 "GenAI Divide" study that 95% of organizations saw no measurable P&L return from generative AI. In our view, most of those projects never named the unit they were trying to move. We wrote more about that pattern in why AI pilots stall before EBIT.

A worked example (hypothetical numbers)

Take a hypothetical mid-size multifamily operator with three acquisitions analysts. Before automation:

• The team receives 60 OMs a week and gives a real first look to about 20.

• Each first-pass underwrite takes about 6 hours, mostly rent roll and T12 entry.

• The team produces 8 first-pass models a week, or 48 analyst-hours.

After automating screening and intake:

• All 60 OMs get a consistent buy-box screen. Analyst review of the screen takes about 5 minutes per deal, or 5 hours a week.

• First-pass time falls to about 2 hours per deal, since the analyst reviews and adjusts instead of typing.

• The same 48 hours now covers the 5 hours of screen review plus about 21 first-pass models.

The team went from 8 to roughly 21 first-pass models a week with no new hire. Producing those 13 extra models the old way would take about 78 more analyst-hours a week, roughly two additional analysts doing nothing but data entry. At a hypothetical fully loaded cost of $150,000 per analyst, that is about $300,000 a year of capacity, and every deal in the inbox now gets a consistent screen. The more important number is harder to model: if one extra deal a year closes because the team saw it in time, the fee and promote on that deal likely exceeds the cost of the whole program. Again, these figures are illustrative, not results from any client.

The metrics to put on the weekly dashboard

• Deals received, deals screened, and deals advanced (screen-to-model conversion)

• Median hours from OM receipt to first-pass model

• Analyst overrides of the screening recommendation (should fall over time)

• Number of material discrepancies flagged per deal (pro forma vs T12)

• Bids submitted and hit rate, by market and broker

How do you start: a 90-day plan for automating acquisitions

Start with one asset type, one stage, and one countable unit. A realistic 90-day plan gets screening and intake running on live deal flow, with analysts using it every day.

Days 1 to 30: define and connect

1. Write the buy box down as explicit rules, and have two people screen the same ten past OMs to find where they disagree.

2. Pick the countable unit and record a two-week baseline (deals screened per analyst per week is the easiest).

3. Gather the last two to three years of models, IC memos, and pass reasons, even if they are scattered across folders.

4. Connect the acquisitions inbox, the deal folder structure, and the property management system for owned-asset actuals.

5. Map the standard model's line items so extracted T12 data has a place to land.

Days 31 to 60: run in parallel

1. Turn on automated screening for all inbound deals, with analysts still doing their own read on a sample.

2. Compare agent recommendations to analyst calls. Track every disagreement and why.

3. Turn on OM, rent roll and T12 intake for deals that advance, populating the first-pass model.

4. Hold a weekly 30-minute review of misses: wrong extraction, wrong mapping, wrong screen.

Days 61 to 90: make it the default

1. Make the agent's screening card the standard input to the pipeline meeting.

2. Add first-draft IC memos for deals going to committee.

3. Compare the countable unit to baseline and report it to the principals.

4. Decide the next stage to add (comps, land sourcing, or pipeline automation) based on where analysts still spend the most time.

The single biggest risk in this plan is not technical. It is adoption. A system the analysts route around is a system you paid for and did not use. Built is not adopted. Put the output where the team already works (the inbox, the model, the pipeline meeting) instead of in a new app they have to remember to open.

Should you build or buy AI for CRE acquisitions?

Buy point tools if you only need extraction from OMs, rent rolls and T12s into a spreadsheet. Build or partner if you want screening and underwriting that reflect your own deal history, house assumptions and portfolio actuals, because no off-the-shelf tool ships with that knowledge.

A firm with an in-house data engineer and a clean deal archive can build a respectable screening pipeline with commercial AI models and some patience. Most acquisitions teams do not have that person, and the real cost is not the first build but keeping it working as your model template, your buy box, and your systems change. The honest competitor for most firms is neither a vendor nor an internal build. It is the status quo: three analysts, a shared drive, and a lot of Mondays like the one at the top of this post. For a fuller breakdown, see buy vs build for the AI context layer.

Frequently asked questions

What is the best AI for commercial real estate acquisitions?

The best choice depends on the stage you are automating. Document extraction tools handle OMs, rent rolls and T12s well. Deal management platforms such as Dealpath handle pipeline tracking. For screening and underwriting that reflect your own history and assumptions, you need a system connected to your deal archive and portfolio data, rather than a standalone extraction tool.

Can AI underwrite a commercial real estate deal on its own?

AI can build a reliable first-pass underwrite from the rent roll and T12 using your house assumptions, and it can flag where the broker's pro forma departs from actuals. Pricing, rent growth, exit cap and the go/no-go remain human decisions, and should be.

What is the difference between a rent roll and a T12?

A rent roll is a snapshot of every unit or suite today: tenant, lease dates, in-place rent and status. A T12 is the trailing twelve months of actual income and expenses. Underwriters reconcile the two to test whether the property collects what the rent roll says it should.

Can ChatGPT or Claude read an offering memorandum?

Yes. General AI assistants summarize OMs well. They cannot see your model, past deals, or portfolio actuals, so their analysis reflects the broker's narrative plus generic assumptions, and there is no audit trail. Use them for a quick read, not as your underwriting system.

How long does it take to automate deal screening?

A focused team can have automated screening running on live deal flow in 30 to 60 days for one asset type, assuming the buy box is written down and past deal files are accessible. Getting analysts to trust and use it daily usually takes the rest of a quarter.

How do I check floodplain and wetlands for a parcel before making an offer?

Start with FEMA's National Flood Hazard Layer for effective flood zones and the U.S. Fish and Wildlife Service's National Wetlands Inventory for mapped wetlands. Both can be checked by address or parcel and automated across a county. Treat both as screening layers, and confirm with a survey and an environmental consultant before closing.

Will AI replace acquisitions analysts?

In our view, no. It replaces the retyping. Analysts who spent half their week entering rent rolls spend it on site visits, broker calls, and pressure-testing assumptions. Teams typically use the freed capacity to look at more deals, not to cut headcount.

What does AI for CRE acquisitions cost, and what is the ROI?

Costs vary widely, from per-seat extraction tools to multi-system implementations. Measure ROI against one countable unit, such as deals screened per analyst per week or hours from OM to first-pass model. If you cannot name that unit before you buy, you will not be able to prove the return after.

Sources

• Dealpath, "The State of AI Readiness in Commercial Real Estate: Momentum Meets Readiness Gap" (2025 survey of 100 institutional investors), click.dealpath.com

• JLL, "Real estate's AI reality check: 90% of companies piloting, only 5% achieved all AI goals" (2025 Global Real Estate Technology Survey, October 28, 2025), jll.com

• Virtualization Review, "MIT Report Finds Most AI Business Investments Fail, Reveals 'GenAI Divide'" (on MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025"), virtualizationreview.com

• Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk" (February 26, 2025), gartner.com

• V7 Labs, "Best real estate underwriting software" (August 2026), v7labs.com

• FEMA Emergency Management Institute, IS-273 course material, "National Flood Hazard Layer (NFHL)," emilms.fema.gov

• U.S. Fish and Wildlife Service, "National Wetlands Inventory," fws.gov

OutcomeCatalyst connects the systems you already run into a governed intelligence layer your team and your agents can reason over. Demos on this site use fictional data. To see this on your own portfolio, start a conversation.