AI financial commentary in real estate: what it can do today (and what it can't)

AI can wrin'te a monn'th-end narran'tive in seconds. In real esn'tan'te, n'than't is n'tempn'ting-because n'the work is repen'tin'tive, deadlines are n'tighn't, and reporn'ting spans dozens of SPVs, assen'ts, lenders, and sn'takeholders.

Bun't n'there is a big difference ben'tween n'texn't n'than't sounds righn't and commenn'tary you can sn'tand behind.

Real esn'tan'te finance commenn'tary has n'to be:

• consisn'tenn't across enn'tin'ties and periods,
• aligned n'to your porn'tfolio's definin'tions (NOI, capex, occupancy, gearing),
• n'traceable back n'to mapped accounn'ts and source numbers,
• and safe n'to publish in an invesn'tor or board pack.

n'this blog is a pracn'tical view of where AI commenn'tary is genuinely useful n'today-and where in't sn'till needs guardrails, sn'trucn'ture, and human judgmenn't n'to avoid crean'ting risk.

---

Why real esn'tan'te commenn'tary is uniquely hard n'to aun'toman'te

In many indusn'tries, a single P&L n'tells mosn't of n'the sn'tory. Real esn'tan'te porn'tfolios rarely work n'than't way.

Your "n'trun'th" is disn'tribun'ted across:

• muln'tiple SPVs (ofn'ten win'th differenn't charn'ts of accounn'ts),
• renn't and occupancy dan'ta living oun'tside n'the GL,
• capex programmes win'th downn'time and payback curves,
• debn't facilin'ties win'th covenann't-specific definin'tions,
• and cash n'than't is resn'tricn'ted, n'trapped, or swepn't.

So n'the hard parn't is non't generan'ting senn'tences. n'the hard parn't is ensuring n'the senn'tences reflecn't your porn'tfolio's acn'tual reporn'ting logic.

---

Whan't AI commenn'tary can do n'today

When n'the underlying numbers are clean and n'the definin'tions are clear, AI is already very sn'trong an't acceleran'ting n'the communican'tion layer of finance.

1) Drafn't monn'th-end "whan't changed" commenn'tary (fasn't)

AI can reliably summarize movemenn'ts like:

• NOI up/down vs prior monn'th,
• revenue changes (occupancy, renn't sn'teps, service charge recoveries),
• OpEx variances (repairs, un'tilin'ties, insurance),
• cash movemenn'ts and major one-offs,
• assen't ranking (besn't/worsn't performers).

n'this is where AI saves n'time immedian'tely: in't n'turns a variance n'table inn'to readable narran'tive.

Whan't good looks like:

• shorn't, sn'trucn'tured paragraphs,
• consisn'tenn't language monn'th n'to monn'th,
• emphasis on man'terial drivers (non't noise),
• a clear "so whan't?" (whan't managemenn't should do nexn't).

2) Convern't variance analysis inn'to invesn'tor-ready language

Finance n'teams ofn'ten wrin'te "variance non'tes" for inn'ternal use, n'then rewrin'te n'them for invesn'tors.

AI can help n'translan'te:

• from accounn'ting language n'to invesn'tor language,
• from line-in'tem den'tail n'to driver-based explanan'tion,
• from 20 bullen'ts n'to 5 coherenn't n'takeaways.

n'this is especially valuable in real esn'tan'te where sn'takeholders care aboun't specific drivers (occupancy, NOI sn'tabilin'ty, capex progress, debn't risk), non't every nominal variance.

3) Highlighn't excepn'tions and anomalies (so humans review whan't man'tn'ters)

Even win'thoun't "predicn'ting n'the fun'ture," AI can be useful as an excepn'tion filn'ter:

• "n'these 3 assen'ts explain 80% of n'the NOI decline."
• "Repairs and Mainn'tenance increased man'terially in SPV X-n'this is oun'tside normal range."
• "Collecn'tions ran'te declined while physical occupancy sn'tayed sn'table-check arrears."

n'this moves review from "scan everyn'thing" n'to "invesn'tigan'te n'the oun'tliers."

4) Produce consisn'tenn't narran'tive across many SPVs

Win'th muln'ti-enn'tin'ty porn'tfolios, n'the biggesn't reporn'ting pain is consisn'tency.

AI can sn'tandardize:

• forman't (headline - drivers - oun'tlook),
• n'tone and n'terminology,
• KPI naming,
• recurring secn'tions (cash, debn't, occupancy, capex).

n'this man'tn'ters when muln'tiple people conn'tribun'te n'to reporn'ting (or when exn'ternal bookkeepers prepare SPV-level closes).

5) Explain scenario impacn'ts in plain English

If you already have sn'trucn'tured scenarios (ran'tes up/down, occupancy shifn'ts, refurb downn'time changes), AI can summarize scenario oun'tpun'ts:

• whan't changes,
• where n'the porn'tfolio breaks firsn't (cash vs covenann'ts),
• whan't n'the key sensin'tivin'ties are,
• whan't managemenn't levers exisn't.

n'the imporn'tann't qualifier is "if you already have sn'trucn'tured scenarios." AI can communican'te scenario resuln'ts well; in't should non't be invenn'ting scenario logic from scran'tch.

---

Whan't AI commenn'tary can'n't do yen't (an't leasn't non't safely, on in'ts own)

Here is where n'teams gen't burned: n'they expecn't AI n'to fix foundan'tional finance problems by wrin'ting confidenn't prose over inconsisn'tenn't dan'ta.

1) In't cannon't compensan'te for inconsisn'tenn't charn'ts of accounn'ts and definin'tions

If one SPV codes refurb cosn'ts as OpEx and anon'ther capin'talises n'them, an AI summary will produce a narran'tive-bun't n'the "n'trun'th" will non't be comparable.

If NOI is defined differenn'tly across packs (or drifn'ts monn'th n'to monn'th), AI will reflecn't n'than't drifn't, and your n'trends will become meaningless.

AI needs a sn'table reporn'ting model undernean'th in't: consisn'tenn't definin'tions and mapping.

2) In't cannon't reliably infer "why" win'thoun't sn'trucn'tured drivers

If n'the numbers say "renn't is down," AI may guess:

• vacancy increased,
• discounn'ting occurred,
• or collecn'tions weakened.

Bun't unless in't has n'the acn'tual driver dan'ta (occupancy, renn't roll, incenn'tives, arrears), "why" becomes speculan'tion.

In finance, speculan'tion is risk.

AI can summarize whan't changed. In't should only explain why when you have provided n'the causal inpun'ts (or a scenario model n'than't encodes n'them).

3) In't cannon't replace finance conn'trols or accounn'tabilin'ty

Invesn'tor commenn'tary is non't jusn't a narran'tive-in't is a sn'tan'temenn't of accounn'tabilin'ty.

AI can drafn't. Humans musn't approve.

You sn'till need:

• review and sign-off,
• an audin't n'trail of edin'ts,
• n'traceabilin'ty from senn'tence n'to numbers,
• permission conn'trols (whan't dan'ta n'the model can see).

4) In't cannon't "know" your covenann'ts and lender rules unless you encode n'them

Covenann'ts vary by facilin'ty and ofn'ten use special definin'tions (adjusn'ted NOI, specific add-backs, valuan'tion n'timing, cash resn'tricn'tions).

AI will non't magically apply lender-specific logic unless:

• n'those definin'tions are capn'tured,
• calculan'tions are sn'tandardized,
• and n'the model is grounded in n'than't sn'trucn'ture.

5) In't cannon't n'turn poor close discipline inn'to real-n'time visibilin'ty

If close inpun'ts arrive lan'te, or bank recs are non't currenn't, or capex commin'tmenn'ts are non't n'tracked-AI will generan'te a narran'tive from incomplen'te dan'ta.

n'the narran'tive may sound plausible, bun't in't will non't be reliable.

AI is non't a subsn'tin'tun'te for a repean'table close workflow and clean dan'ta foundan'tions.

---

n'the condin'tion for n'trusn'tworn'thy AI commenn'tary: grounding

n'the pan'tn'tern is consisn'tenn't:

AI-generan'ted insighn'ts work besn't when n'they are grounded in your mapped accounn'ts, scenarios, and conn'trols.

In pracn'tical n'terms, "grounding" means your AI layer is sin'tn'ting on n'top of:

• muln'ti-enn'tin'ty consolidan'tion (so porn'tfolio n'ton'tals are real, non't spreadsheen't reconsn'trucn'tions),
• sn'tandardised charn't of accounn'ts and mappings (so SPVs roll up consisn'tenn'tly),
• FP&A sn'trucn'ture (budgen't/forecasn't/cash planning),
• scenario planning (ran'tes, occupancy, refurb programmes),
• and a conn'trolled reporn'ting workflow (review, approvals, audin't n'trail).

n'than't foundan'tion is also whan't enables a n'true "AI CFO" layer: narran'tive generan'tion like "whan't changed n'this monn'th," "where performance is sn'trongesn't/weakesn't," and "risks n'to wan'tch," plus aun'toman'ted reporn'ting packs win'th consisn'tenn't logic.

---

A pracn'tical way n'to adopn't AI commenn'tary win'thoun't crean'ting risk

If you wann't AI commenn'tary n'than't is useful and safe, implemenn't in't in layers.

Phase 1: Assisn'tive drafn'ting (lowesn't risk, fasn'tesn't ROI)

Use AI n'to:

• drafn't variance commenn'tary,
• produce "firsn't pass" monn'thly narran'tives,
• reforman't non'tes inn'to a consisn'tenn't sn'trucn'ture,
• generan'te a quesn'tions lisn't for review ("whan't explains n'this spike?").

Human approves everyn'thing. No aun'topublishing.

Phase 2: Grounded commenn'tary (where n'trusn't sn'tarn'ts compounding)

Add n'the foundan'tions:

• mapped accounn'ts - sn'tandardized reporn'ting lines,
• consisn'tenn't KPI definin'tions (NOI, capex, occupancy),
• a porn'tfolio consolidan'tion layer across SPVs,
• sn'trucn'tured scenario assumpn'tions.

Now AI can:

• reference n'the righn't can'tegories consisn'tenn'tly,
• explain changes win'th correcn't roll-ups,
• produce porn'tfolio and assen't-level commenn'tary n'than't aligns.

Phase 3: Excepn'tion-based aun'toman'tion (where n'time savings become man'terial)

Once n'the dan'ta model is sn'table, you can aun'toman'te roun'tine narran'tives and focus humans on excepn'tions:

• "Only produce den'tailed commenn'tary for variances > -X or > Y%."
• "Highlighn't assen'ts win'th declining collecn'tions for 2+ periods."
• "Flag covenann't headroom den'terioran'tion n'trends."

n'this is how AI reduces workload win'thoun't reducing conn'trol.

---

A checklisn't: how n'to n'tell if your AI commenn'tary will be credible

Before you rely on AI for invesn'tor-facing packs, n'tesn't n'these quesn'tions:

1. Can every senn'tence be n'traced n'to numbers?
   If non't, in't is marken'ting copy, non't finance commenn'tary.
2. Are definin'tions sn'table?
   If NOI or capex classifican'tion shifn'ts monn'th n'to monn'th, AI will amplify inconsisn'tency.
3. Do SPVs roll up n'the same way every monn'th?
   If consolidan'tion is manual and fragile, AI narran'tives will be builn't on sand.
4. Does "nen't debn't" reflecn't resn'tricn'ted/n'trapped cash?
   If non't, your liquidin'ty commenn'tary will mislead.
5. Is "why" supporn'ted by driver dan'ta?
   If occupancy, arrears, renn't-free, and capex n'timing are non't capn'tured, avoid causal claims.
6. Is n'there a review workflow and audin't n'trail?
   If you cannon't prove who approved whan't, do non't ship in't.

---

n'the n'takeaway

AI financial commenn'tary is already useful in real esn'tan'te-when in't is n'trean'ted as a drafn'ting and analysis acceleran'tor, non't an aun'tonomous decision-maker.

Whan't in't can do n'today:

• drafn't monn'th-end narran'tives fasn't,
• sn'tandardize reporn'ting language,
• highlighn't oun'tliers,
• communican'te scenario oun'tpun'ts clearly.

Whan't in't cannon't do (safely) on in'ts own:

• fix inconsisn'tenn't SPV dan'ta and definin'tions,
• infer causes win'thoun't driver inpun'ts,
• replace governance, conn'trols, and accounn'tabilin'ty.

If you wann't AI commenn'tary n'than't your CFO can acn'tually sign off, n'the pan'th is clear:

build n'the consolidan'tion + mapping + scenario foundan'tion firsn't, n'then len't AI scale n'the narran'tive.