Accounting AI agent · Client requests
Client Document-Collection & Follow-Up AI Agent
Build the request list, ask the client, check what arrives against the request and report what is still outstanding — inside approved wording, contact preferences, chase limits and human escalation.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Human review
- At escalation
- After launch
- Agent Care
What this agent does
Automates the asking, checking and chasing layer
In
- 01
Build the request list from the engagement scope, the prior-year file and the entity structure.
- 02
Take documents in from the client portal, email intake, secure upload or messaging channel.
Reason
- 03
Classify each arriving document by type, entity, period and the request it answers.
- 04
Check it against that request — right entity, right period, complete and readable.
- 05
Reconcile against what is already on file so nothing already supplied is asked for again.
Decide
- 06
Detect partial, mismatched or unreadable uploads and requests no channel has satisfied.
- 07
Follow the approved chase cadence and contact preferences, and hold anything else for a person.
Out
- 08
Report outstanding items, ageing and chase history to the engagement owner and the partner.
- 09
Message clients and write to the engagement file only inside the approval boundaries agreed during implementation.
Product statement
The agent asks, receives, checks and reports; deciding that a client has complied, and every message outside the approved set, stay with a person.
Example workflow
One missing document, end to end
- 1Request list assembledEngagement scope, prior-year file, entity structure and the firm's standard list
- 2Request issuedApproved wording, named contact, agreed channel and due date
- 3Document receivedPortal upload, email attachment, secure link or forwarded message
- 4Checked against the requestDocument type, entity, period, legibility, completeness and what is already held
No human action required
Stages 1 to 4 run without a person in the loop — the first ask and each on-cadence chase use wording the firm has already signed, so the lane stays empty until the gate.
Right entity, right period, readable and complete — follows the approved path.
Wrong period, wrong entity, partial or unreadable — enters human review.
Human review
The upload is held with the request it was matched to, the checks it failed and the chase history.
Approve · Correct · Request review
Approved — handed back
- 6Filed and status updatedOnly where write access and approval policy allow it; the request is not closed by the agent
- 7Outcome evaluatedCompletion, days to complete, chases per response, mismatches and opt-outs
Overrides
Rejections and re-matches made in review are counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Deciding a request has been satisfied.
- Closing the request list.
- Any message outside the approved wording.
- Escalating to a partner.
Automation boundary
Agent acts unaided
- ✓Build the request list from the engagement and the prior-year file.
- ✓Issue approved requests and on-cadence chases to named contacts.
- ✓Classify arriving documents and check them against the request.
- ✓Report outstanding items, ageing and chase history to the owner.
Client messages and write actions run only inside the approval boundaries agreed during implementation. Closing a request is not one of them.
- Chasing beyond the agreed cadence.
- Contacting anyone not on the engagement.
- Handling a document sent to us in error.
- Changing cadence, wording or preferences.
Example output
One request and one upload, annotated
Everything the agent proposes is attached to the request it claims to answer.
Collection output · single request
Illustrative example- Request
- Year-end bank statements
- Channel
- Client portal upload
- Received
- Two PDFs
- Matched to
- Bank statements request
- Confidence
- 93%
- Check result
- Period incomplete — held
As received
Taken from the upload and the channel it arrived on — nothing on this side is inferred, and no file is opened outside the engagement.
Evidence used
- Request-list entry
- Period and entity read
- Prior-year document set
Why it is held
Part of the requested period is missing, so the request stays open and the client is told what is still needed.
What the score decides
Below the configured threshold the upload routes to review instead of the approved path. Marking the request satisfied is a human action either way.
Action
ApproveCorrectRequest reviewValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
All open requests
Across every engagement
Apply client-specific context
Use the engagement scope, prior-year file, entity structure and the firm's approved wording.
Remove the routine chasing
Requests and on-cadence reminders go out on approved wording without anyone drafting them.
Put people on the awkward cases
Partial, mismatched and disputed uploads move to review instead of every response being opened by hand.
Build an evidence trail
Retain every message sent, document received, check result, confidence, evaluator result and human correction — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Client portals
- Karbon
- Canopy
- TaxDome
- Liscio
- Suralink
Document management
- SharePoint
- iManage
- NetDocuments
- Virtual Cabinet
Practice & engagement
- Job records
- engagement files
- Caseware
- workpapers
Agent
Client document collection & follow-up
- Issues requests
- Checks documents
- Reports outstanding
Messaging & signature
- Email intake
- SMS
- DocuSign
- Adobe Acrobat Sign
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and your client
Each control wraps the one inside it. A message clears every layer before a client sees it, and closing a request sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether the draft may stand
- L3
- Decides how often a client may be contacted
- L4 – L5
- Keep sign-off human and the record complete
- L6
- Pulls outreach back when signals degrade
How Nestack evaluates it
Evaluate the full workflow — not only the final match.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the message the client receives
- E1Final-output evaluationDid the document actually answer the request?
- E2Step-level evaluationDid the agent read the right entity, period and document type?
- E3Tool evaluationDid it write to the correct engagement file and request line?
- E4Confidence calibrationDo low-confidence matches actually contain more errors?
- E5Slice evaluationHow does performance change across specific client cohorts?
- E6Business outcomeHow many requests closed, and what did each response cost in chases?
Floor — the complete file the engagement needs
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle.
Agent lifecycle
Direction of processing →
01 · Request build
2 modes
- DC-01
Chase for a supplied item
The document arrived on another channel and was never linked.
- DC-02
Unreadable upload passes
A photographed statement is filed that nobody can read.
Stage gathers
Engagement scope, prior-year file and entity structure
02 · Match & check
2 modes
- DC-03
Wrong period or entity
A prior-year or sister-entity file is accepted.
- DC-04
Partial upload passes
Most of the period arrives and it reads as met.
Stage proposes
Type, entity, period, completeness and confidence
03 · Outreach / write
1 mode
- DC-05
Over-chasing past the cadence
Parallel workflows chase one client past the limit.
Stage writes
Only where wording, cadence and write access allow it
04 · Output
1 mode
- DC-06
Chase to the wrong contact
A former bookkeeper or unrelated director gets it.
Stage returns
The message the client and the engagement owner see
05 · Change / Version
1 mode
- DC-07
Silent matching regression
A wording change lowers match quality.
Stage tracks
Model, prompt, wording and cadence changes
- Sev-1 · reaches the client
- Sev-2 · wrong document enters the file
- Sev-3 · input degrades, request routes to review
Affected slices
A high completion rate can hide concentrated risk
Aggregate completion can look acceptable while a small number of client cohorts absorb most of the chasing and nearly all of the mismatched uploads. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| First-year and newly onboarded clients | 5.4% | 3.4× | Review | |
| Multi-entity client groups | 4.0% | 2.5× | Review | |
| Photographed and scanned uploads | 3.0% | 1.9× | Watch | |
| Portal-active repeat clients | 1.1% | 0.7× | Normal |
Bar: lift vs. portal-active baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
The client's reaction is evidence too
This is the agent your clients hear from, so its signals come from them — an opt-out, a rejected upload, a chase that produced nothing — and every wording change is re-approved first.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Chases per response, opt-outs or rejected uploads rise in a cohort.
- 02
Diagnose
Traced to the request wording, the channel, the contact record or the check.
- 03
Improve
The wording, cadence or check is re-approved by the firm and version-linked.
- 04
Verify
Re-run against held-out requests and the messages they would have produced.
- 05
Learn
The objection becomes a standing rule in the approved wording library.
Learn → Detect
The return edge. Each re-approved wording and cadence limit binds the next cycle before a client sees it.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, intake, matching and messaging, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Request-list templates and rollforward.Week 1, Week 2
- 03Contact records and consent mapping.Week 1, Week 2
- 04Portal, email and upload intake.Week 2
- 05Document classification and request matching.Week 2, Week 3
- 06Entity, period and completeness checks.Week 3
- 07Message wording and firm tone review.Week 3, Week 4
- 08Chase cadence, limits and opt-out rules.Week 3, Week 4
- 09Human review and escalation workflow.Week 4
- 10Evaluation suite and regression cases.Week 4, Week 5
- 11Engagement filing and status reporting.Week 5
- 12Observability, deployment and Agent Care handover.Week 5, Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne engagement type | ProductionProduction integration | AdvancedMultiple entities / systems |
|---|---|---|---|
| Introduced at Pilot | |||
| Request lists and document checks | ✓ | ✓ | ✓ |
| Human approval | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Approved message wording | — | ✓ | ✓ |
| Chase cadence and rate limits | — | ✓ | ✓ |
| Opt-out and preference handling | — | ✓ | ✓ |
| Portal and email intake | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-entity request routing | — | — | ✓ |
| Multi-stage approvals | — | — | ✓ |
| Enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on engagement types, portal and document-management integrations, client volume, message-approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your standard request lists and prior-year engagement files
→Request-list templates and prior-year rollforward
Week 102Client and contact records, and who may be contacted about what
→Contact records, permissions and consent mapping
Week 103Access to the portal, mailbox and document management system
→Portal, email and upload intake, then integration setup
Week 204A representative sample of what clients actually send you
→Document classification, request matching and the check baseline
Week 205Your approved wording, chase rules and opt-out list
→Message wording, chase cadence, rate limits and opt-out handling
Week 306The requests you chased hardest last year
→Evaluation suite, regression cases and failure-mode testing
Week 407Named reviewers and the partner who owns escalation
→Human review and escalation workflow, then pilot and validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 carries evaluation and the first supervised sends to a real client.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Workflow discovery, contact and consent mapping, automation boundary
W2Intake channels, classification and matching baseline
W3Message wording, chase cadence and rate limits
W4Evaluation suite, escalation workflow and dry-run testing
W5Supervised sending to one engagement, every message read
W6Full collection cycle on a live engagement, then handover
- Reading the band
- Bars span only the weeks their work is named in. No message reaches a client before week 5.
- At the end of W6
- One collection cycle has run under review, chase log and opt-outs checked, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Accounting AI agent
Build a document-collection agent around your request process.
Show us your request lists, your portal and mailbox, and the wording and chase cadence you would defend to a client. We'll agree the wording, the cadence limits and the point where a person takes the conversation over.
- Typical delivery
- 4–6 weeks
- Deployment
- Your practice stack
- Human review
- At escalation and off-template
- After launch
- Agent Care available
Nestack Agents · Client document collection & follow-up
AGT-ACC-10 · Agent Care available after launch