Financial-statement OCR
Multi-format extraction puts portfolio-company and sub-fund reporting into one definitional table, with the source page kept beside every field.
- Multi-format capture
- Field-level sampling
Capability · AI & risk
Put agents where answers can be checked, and accountability stays with people
The company reports the first AI-agent investment-management platform in the sector. The judgement was scene selection: extraction, indicator assessment, behaviour logging and effort costing have objectively checkable answers, so benefit is measurable on day one. "The AI decides the investment" cannot be verified and should not be sold. Agents sit inside existing nodes and emit structured fields.
Position
Whether an answer can be checked today decides which AI scenes we take
Does the extracted figure match the source document? Did the indicator fire as defined? Is the sensitive operation on record? Was the hour captured? These four have objective answers, so one run quantifies both the benefit and the error. "Should we back this company" has no same-day answer; putting a model there dresses an unverified judgement up as a system output.
Our agents are therefore workflow nodes, not a chat box. Input is existing evidence; output is a structured field with its source document and confidence attached, and a person signs the conclusion at the end of the node.
Where we draw the line
Why we do not build "AI that decides investments"
Two reasons, neither of them technical. The first is verifiability: an investment decision is proved right or wrong at exit, years later, while the model must produce a confident-looking score today. That feedback gap supports no acceptance test and demonstrates no superiority over a committee. The second is accountability: investment decisions are answered for by named committee members under a written policy. Handing the judgement to a model removes the accountable party, and in a dispute the institution cannot produce a traceable basis for what was decided.
What we compress is evidence gathering, reconciliation and costing — the work that already consumes most of a committee member’s preparation. Judgement, veto and signature remain human; the system’s job is that everyone sees the same defined facts.
Three engineering constraints we do not negotiate
These appear in the contract and in the acceptance method, not in product adjectives.
- Every model output keeps its source evidence and confidence
- Any agent can be disabled by client permission; the flow still runs
- Alerts trace back to the indicator definition that fired
- On intranet, models run inside the client environment with no egress
- Sample sets and pass criteria are confirmed by the business beforehand
Five steps, ending in an agent you can switch off
The order matters: before the business signs a sample set, no accuracy figure means anything.
Screen scenes for objectively checkable answers
Work through the break inventory and keep only steps where right and wrong can be judged the same day. Output: scene list, plus written reasons for what we declined.
Build the sample set and the pass rule
The business labels samples from real evidence and agrees field-level pass criteria and the sampling ratio. Output: frozen, signed benchmark set.
Run the model where the data already is
Public cloud, private cloud or intranet each have their own connection pattern; on intranet the model sits inside your network boundary. Output: topology note and egress confirmation.
Sample-test and publish the error list
Test against the benchmark, list every misjudgement with its cause, and mark which errors need human review and which may pass. Output: sampled-accuracy report and error register.
Wire the kill switch and prove the fallback
Each agent has an independent switch; with it off, the node reverts to manual work and the process still closes. Output: fallback regression record, demonstrated live at acceptance.
Artefacts, duration, acceptance
4–10 weeks is the standard range, counted from agreement of the work-breakdown structure; scene count and document-format spread set the position in the range. We do not promise every agent in production inside two weeks.
| Artefact | Typical duration | Acceptance method |
|---|---|---|
| Scene list and declined-scope note | Within 4–10 weeks | Business confirms an objective test exists for each retained scene |
| Labelled benchmark set and pass rule | Within 4–10 weeks | Signed by the business, then frozen |
| Statement OCR field extraction | Within 4–10 weeks | Field-level sampling against source documents; misses enter the error register |
| Risk indicator alerts and judgement log | Within 4–10 weeks | 18 indicators recalculated to definition; triggers and confirmations traceable |
| Desktop behaviour evidence trail | Within 4–10 weeks | Sampled replay of operation chains, exportable |
| Effort capture and costing basis | Within 4–10 weeks | Run in parallel with manual records for one cycle |
| Agent kill switch and regression record | Before go-live | An agent is disabled on site; the workflow still closes |
Definition source
Agents judge definitions, not private opinions
Every warning indicator points at a field and a formula in the data dictionary. Changes travel through the governance process, so a model never grows a second private definition — the precondition for risk and audit accepting an AI conclusion.
- Indicators map to dictionary fields
- Changes re-run the benchmark set
- Formula version stored with the result
Four production scenes
We ship these four, matching public disclosure
Scene names follow the published AI scope. Each one sits inside an existing step and emits reviewable structured fields.
Risk-model judgement
Custom indicators, with the 18 warning metrics recalculated daily and every trigger and confirmation logged.
- Custom indicators
- 18 warning metrics
Desktop behaviour control
Monitoring of sensitive operations gives lending and quasi-financial work a replayable compliance record; rules and visibility stay yours.
- Sensitive-operation alerts
- Replayable chain
Effort accounting
Automatic capture of hours by project makes diligence and committee effort enter a costable definition for the first time.
- Automatic capture
- Cost roll-up
Where agents attach
Agents hang off existing nodes rather than adding a parallel process, and follow the hosting mode chosen for the platform.
DealFlow project attachmentsDiligence and committee packsPortfolio company monthly and annual reportsSub-fund reporting templatesPortfolio risk monitoringAI desktop behaviour systemVDR documentsPublic cloud / private cloud / intranet
Evidence
What each scene is measured on afterwards
Statement extraction is measured by sampled accuracy and the error register; indicator judgement by alert confirmation rate; behaviour evidence by whether the operation chain replays in full; effort capture by a parallel cycle against manual records. All four share the definitions used on our client-impact pages.
Acceptance definitionsIn writing
We publish where the line sits
Which steps agents should enter, and which they must not — including why we refuse to hand an investment decision to a model — is argued openly in our insights, so you can quote it in a scope review rather than take it on trust.
Read the articleWho should not buy this
Anyone wanting a model to say whether to invest: we will not build it and will record the refusal in the scope review. Anyone planning to replace a review post with an agent: agents shorten evidence work, and the signature still needs a role behind it.
Institutions whose source documents are not yet digital — and not yet in a dictionary — should take data governance first. OCR cannot read a document that does not exist, and an indicator cannot be calculated from an undefined field.
Questions asked before agents go live
Answers match our acceptance method and can be quoted in an internal note.
Will the AI agents make our investment decisions?
No. Scope stops at evidence gathering, reconciliation, indicator recalculation and logging. Committee members and written policy carry the decision, because its correctness cannot be tested on the day and accountability cannot be delegated to software.
On an intranet deployment, does data leave?
No. OCR and indicator models run inside your environment with no outbound path, and the topology and egress position are written into the acceptance note. Enabling each agent remains your permission.
Who owns a wrong extraction?
Errors surface in the sampled test and the error register; the split between human review and auto-pass is agreed beforehand. Because each output keeps its source document and confidence, any conclusion can be traced back to the evidence and the rule applied.
What happens when we disable an agent?
The node falls back to manual work and the process still closes. Each agent has its own switch, and the fallback regression is demonstrated before go-live.
Can the 18 warning indicators follow our own policy?
Yes. Custom indicator assessment is one of the four scenes: new or amended indicators enter the dictionary with a formula and threshold, the model recalculates to the new definition, and earlier conclusions keep the version in force at the time.
Is effort capture employee surveillance?
It measures project effort and cost, and its scope and visibility levels are set by your permissions. It is a separate module from desktop behaviour control, and each can be enabled or disabled on its own.
Next step
Bring three real documents
One company report, one sub-fund return, one operation log. We judge on the spot which steps have objective answers, which should stay manual, and where the 4–10 week range lands.