Accounting AI agent · Tax
Tax-Question Copilot (Jurisdiction-Aware)
Answer internal tax questions with a drafted response, the jurisdiction and tax year it applies to and every authority cited — held for a qualified reviewer, with evaluation, audit trails and guardrails built in.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Human review
- Before release
- After launch
- Agent Care
What this agent does
Drafts the answer and shows its authority
In
- 01
Ingest tax questions from supported chat, email, ticketing or research-tool channels.
- 02
Normalise each question into jurisdiction, tax year, entity type and the facts supplied.
Reason
- 03
Retrieve statute, regulations, rulings and published guidance in force for that period.
- 04
Apply the firm's precedent library, agreed house positions and the engagement file.
- 05
Draft an answer that separates the rule, its application and what it does not cover.
Decide
- 06
Check every citation exists, still stands and supports the sentence it sits under.
- 07
Route advice-class, unsettled and low-confidence questions to a qualified reviewer.
Out
- 08
Retain the question, retrieved authorities, draft, confidence and reviewer decision.
- 09
Release or write answers only inside the approval boundaries agreed during implementation.
Product statement
The agent drafts and cites inside the approval boundaries agreed during implementation; a qualified person signs the position.
Example workflow
One question, end to end
- 1Question receivedChat, email, ticket or the firm's research tool
- 2Scope resolvedJurisdiction, tax year, entity type and the facts already on the engagement file
- 3Authority retrievedStatute, regulations, rulings, published guidance and firm precedent in force for the period
- 4Controls appliedCitation-existence checks, currency checks, scope guardrails and confidence threshold
No human action required
Stages 1 to 4 run without a person in the loop — nothing has left the agent yet, so the lane stays empty until the gate.
Returns to the requester as cited draft research.
Held for a qualified reviewer.
Reviewer sign-off
The draft is held with its citations, jurisdiction, tax year and confidence.
Approve · Amend · Escalate
Signed — handed back
- 6Answer issued and filedReturned to the requester as draft; issued outside the firm only after the reviewer signs
- 7Outcome evaluatedCitation accuracy, authority currency, reviewer amendments, escalations and business outcome
Amendments
Corrections and escalations 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
- Any answer issued outside the firm.
- Filing positions and disclosure decisions.
- Questions where the authorities genuinely conflict.
- Questions turning on facts not on file.
Automation boundary
Agent acts unaided
- ✓Retrieve authority in force for the jurisdiction and period.
- ✓Draft the answer with its citations and its reasoning.
- ✓Check that cited authority exists, still stands and applies.
- ✓Route advice-class and low-confidence questions to review.
Answers are released and written only inside the approval boundaries agreed during implementation.
- Positions resting on non-precedential written determinations.
- Answers on legislation still inside transition.
- Anything framed as planning or structuring advice.
- Changes to the authority corpus or confidence thresholds.
Example output
One drafted answer, annotated
Everything the agent proposes is attached to the question and the authority it came from.
Draft answer · single question
Illustrative example- Question
- Availability of a new relief
- Entity
- Multi-state operating company
- Tax year
- 2025
- Draft answer
- Available federally; states differ
- Confidence
- 88%
- Authority cited
- Statute, regulation, notices
As received
The question, entity and tax year exactly as the requester gave them — the agent does not assume a jurisdiction that was not stated.
Evidence used
- Statute in force
- Current regulation text
- State conformity notices
Why it splits
The states in scope do not adopt the federal provision for this period, so the two answers are kept apart.
What the score decides
Below the configured threshold the draft is held for a qualified reviewer instead of returning to the requester.
Action
ApproveAmendEscalateValue
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 tax questions
From the firm's channels
Apply the firm's own positions
Use the precedent library, agreed house positions and the facts already on the engagement file.
Cut the routine research time
Settled, in-scope questions come back drafted and cited instead of being researched from scratch.
Put reviewers on the hard questions
Advice-class, unsettled and low-confidence questions reach a qualified reviewer instead of every question doing so.
Build an evidence trail
Retain the question, retrieved authorities, draft, confidence, evaluator result and reviewer decision — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Tax research platforms
- Checkpoint
- CCH AnswerConnect
- Bloomberg Tax
- IBFD
Primary sources
- Legislation
- regulations
- rulings
- Published guidance
- authority feeds
Firm precedent
- Prior memos
- house positions
- Document management
- engagement files
Agent
Tax-question copilot
- Reads authority
- Drafts and cites
- Routes to review
Practice workflow
- chat
- ticketing
- Practice management
- approval workflow
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 a signed answer
Each control wraps the one inside it. A draft clears every layer before anyone is asked to rely on it.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether the answer may stand
- L3
- Decides whether a qualified person signs it
- L4 – L5
- Keep the record and the authority current
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the full workflow — not only the final answer.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the answer the reviewer sees
- E1Final-output evaluationWas the answer correct for that jurisdiction, year and entity type?
- E2Citation evaluationDoes every cited authority exist, apply and still stand?
- E3Step-level evaluationDid it retrieve and weigh the right authority for the period?
- E4Escalation calibrationAre unsettled and advice-class questions actually held back?
- E5Slice evaluationHow does performance change across specific question cohorts?
- E6Business outcomeHow many answers were amended or escalated in review?
Floor — the outcome the client pays for
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 · Retrieval
2 modes
- TQ-01
Superseded authority cited
A revoked, obsoleted or modified ruling is treated as current.
- TQ-02
Authority coverage gap
The corpus holds nothing for the jurisdiction or period asked.
Stage gathers
Statute, regulations, rulings, guidance and firm precedent
02 · Reasoning
2 modes
- TQ-03
Fabricated citation
A section, ruling or case that does not exist is quoted.
- TQ-04
Jurisdiction or period mismatch
The answer is right for another state, country or tax year.
Stage proposes
The answer, the authority behind it and confidence
03 · Tool / write
1 mode
- TQ-05
Advice-boundary crossing
A draft is released as though a qualified person had signed it.
Stage writes
Only where release policy and approval boundaries allow it
04 · Output
1 mode
- TQ-06
Unsettled point answered flat
A genuinely contested position is returned without escalation.
Stage returns
The answer the requester and the engagement file see
05 · Change / Version
1 mode
- TQ-07
Silent corpus drift
An authority update is not reflected in the retrieved set.
Stage tracks
Authority updates, model and prompt changes
- Sev-1 · unsupported answer could be relied on
- Sev-2 · wrong answer for that taxpayer
- Sev-3 · retrieval degrades, routes to review
Affected slices
Overall health can hide concentrated risk
Aggregate answer quality can look acceptable while a small number of question cohorts carry most of the citation and currency failures. 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 |
|---|---|---|---|---|
| Cross-border questions | 5.8% | 3.4× | Review | |
| Recent-legislation questions | 3.9% | 2.3× | Review | |
| Multi-state nexus questions | 2.9% | 1.7× | Watch | |
| Settled federal questions | 1.0% | 0.6× | Normal |
Bar: lift vs. settled-question baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
The loop does not end at Learn
Each cycle leaves the corpus re-tagged, a position re-verified and an escalation rule in force — that is what the next detection is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Citation or currency failures rise in a question cohort.
- 02
Diagnose
Failure isolated to retrieval, authority weight, prompt or scope.
- 03
Improve
Corpus re-tagged or an escalation rule added — approved and version-linked.
- 04
Verify
Affected questions are re-asked against the updated corpus.
- 05
Learn
The question becomes a standing test; the signed position joins firm precedent.
Learn → Detect
The return edge. The next cycle starts against a re-tagged corpus and one more escalation rule.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, authority corpus, agent workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and advice-boundary definition.Week 1
- 02Source-system and research-platform assessment.Week 1, Week 2
- 03Jurisdiction, tax and question-class scoping.Week 1, Week 2
- 04Authority corpus and effective-date tagging.Week 2
- 05Retrieval, citation checking and drafting.Week 2, Week 3
- 06Confidence scoring and escalation routing.Week 3
- 07Reviewer sign-off workflow.Week 3, Week 4
- 08Research and document-system integration.Week 5
- 09Evaluation suite and regression cases.Week 4, Week 5
- 10Guardrails and advice-boundary controls.Week 4
- 11Observability and trace instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.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 jurisdiction | ProductionProduction integration | AdvancedMultiple jurisdictions |
|---|---|---|---|
| Introduced at Pilot | |||
| Drafted answers with citations | ✓ | ✓ | ✓ |
| Reviewer sign-off | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Firm precedent library | — | ✓ | ✓ |
| Escalation workflow | — | ✓ | ✓ |
| Approved release actions | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-jurisdiction coverage | — | — | ✓ |
| Multi-stage review | — | — | ✓ |
| High query volume | — | — | ✓ |
| Enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on jurisdictions in scope, research sources, question volume, review 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
01The jurisdictions, taxes and question types in scope
→Jurisdiction, tax and question-class scoping
Week 102The advice boundary and who may sign what
→Advice-boundary definition and approval controls
Week 103Access to your research subscriptions and sources
→Source-system and research-platform assessment
Week 204Prior memos, house positions and precedent files
→Authority corpus build, effective-date tagging and precedent indexing
Week 205Confidence and escalation thresholds
→Confidence scoring, escalation routing and guardrails
Week 306Questions your team has found hard to answer
→Evaluation suite, regression cases and failure-mode testing
Week 407Named reviewers or test users
→Reviewer sign-off workflow, then pilot workflow and production 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 both evaluation and launch work.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Workflow discovery, advice boundary and question-class scoping
W2Authority corpus, source access and the research baseline
W3Agent workflow, citation checking and reviewer controls
W4Evaluation suite, advice-boundary guardrails and failure modes
W5Research-platform integration, pilot and targeted corrections
W6Sign-off on live questions, verification and Agent Care handover
- Reading the band
- Bars cover only the weeks their work is named in. Build starts once the advice boundary is agreed in week 1.
- At the end of W6
- Live questions are answered under sign-off, then Agent Care monitors the corpus and the escalations.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Accounting AI agent
Build a tax-question copilot around your research workflow.
Show us the questions your team asks, the sources you already licence and who signs what. We'll map the workflow, set the advice boundary and recommend the safest path to production.
- Typical delivery
- 4–6 weeks
- Deployment
- Your research stack
- Human review
- Before release
- After launch
- Agent Care available
Nestack Agents · Tax-question copilot (jurisdiction-aware)
AGT-ACC-02 · Agent Care available after launch