Stephen ReaProduct, design & AI strategy
AI does the preparation. People make the call.
A clear view of what Thread has prepared, what still needs checking and which decisions remain yours.
An accountant can have the ledger, the documents and a practice system, yet still need to reconstruct what is missing, who owns the next step and whether an approval is current. Paper Thread brings that context together around the work.
The morning overview puts AI-prepared work in front of Maya: a possible invoice match with its source context and a focused evidence request ready for review. Each item shows what Thread did, checks still outstanding and the action that needs a person.
Browse the product screensPrepared by Thread. Checked and authorised by you.01 / 03
A suspected duplicate is a question, not a verdict. The accountant can inspect the source records, obtain the missing statement and record a decision before any correction is made.
The finding, its source and the person who decides.02 / 03
The reviewer sees the current packet, the evidence behind it and what has changed. When the source changes, dependent approvals become superseded and the work returns for review.
Current evidence. Explicit authority. A retained decision.03 / 03
Edit a source record, watch Thread map the new snapshot, import a sample statement and carry the evidence through to a human review.
Try the prototypeFictional data, simulated AI and already-connected Xero and QuickBooks. Edits and CSV files stay in your browser. No provider connection, email, source write or filing takes place. Open full size ↗
The core journey is simple to describe. The edge cases are where the design has to earn trust.
Seven proposed journeys connect the core workflow to the people, evidence and exceptions around it. Each one is a designed scenario to validate.
Senior accountant
A current review packet with traceable evidence and a named reviewer.
Unsupported findings are dismissed with a reason or deferred to an owner; missing evidence remains a blocker.
Accountant and client contact
Verified evidence attached to the correct client and period.
The wrong period or incomplete coverage cannot silently satisfy the request.
Preparer
Traceable staged data, with exclusions still visible.
Wrong entity or period is quarantined; exact duplicates and excluded rows remain accounted for.
Assigned reviewer
A version-bound sign-off with evidence of the actual outcome.
Changed evidence supersedes dependent approval. Export does not mean filed or accepted.
Client manager and compliance owner
The correct client, an agreed service and explicit authority.
Missing acceptance prerequisites prevent activation; restricted compliance evidence stays with its designated owner.
Relationship owner and junior preparer
Continuity without opening the whole client file.
Expiry blocks new retrieval and queued actions; it cannot claim back files already exported.
Mandate owner and connection owner
Inspectable assistance with a controlled recovery path.
A timeout is an unknown outcome. Check what happened before attempting an action again.
These are proposed journeys derived from the requirements and designed screens. They have not yet been observed or tested with accountants.
Assistance earns trust by showing its work: what it used, what it did, what remains uncertain and where a person must intervene.
Suggest possible matches, explain a finding from permitted evidence, draft a focused client request and assemble a review summary. The morning dashboard brings this prepared work to the person who must review it, with the source and the uncertainty available.
Use explicit logic for calculations, parsing, access checks, version dependencies, retry behaviour and action limits. Ordinary automation should do ordinary deterministic work.
The accountant confirms the accounting conclusion, verifies evidence and approves external requests. The assigned reviewer approves the current work. Filing, ledger writes, billing and payments are not autonomous actions in the initial concept.
A mandate states its purpose, scope, owner, allowed actions and expiry. The dashboard exposes that delegation and its remaining approval steps. Activity distinguishes completed preparation, drafted actions, failed steps and pending work. Pausing stops pending work while retaining what already happened.
Next test: Compare assisted and unassisted work for total effort, false alarms, review returns and the user’s ability to explain and stop the process.
Access follows the client, engagement and task relationship, with explicit capabilities and expiry.
A junior may work on bank evidence within an assigned client without seeing payroll or tax advice. Review, export and administrative capabilities are granted separately.
The same limits apply to search, AI context and background tasks. Hiding a field after retrieval is not sufficient. Expired access produces a generic unavailable state without protected client details.
Missing, stale, conflicting, unsupported and not tested are distinct states. Each identifies its owner and resolution route rather than disappearing into a red or green score.
Connecting a source is different from accepting a client, agreeing an engagement or gaining authority to act. These prerequisites have separate owners and evidence.
Next test: Exercise restricted-role and expired-access scenarios, including AI retrieval and pending background actions, before any real client material enters a pilot.
I mapped the wider practice jobs, narrowed the first workflow and designed the screens, states and handoffs that make the proposition inspectable.
The opportunity is not simply a shared database or another dashboard. The first design loop carries a finding through evidence, correction, refresh and current-version review across the firm’s chosen tools.
The dashboard provides orientation; the work views provide depth. I kept the next decision, its owner and its evidence connected, so the user can move from a portfolio-level signal into a specific task.
The screen set includes incomplete imports, source outages, restricted access, expired cover, changed approvals and uncertain external outcomes. These states test whether the product’s promise survives real work.
Warm neutral surfaces, forest navigation, restrained emphasis and explicit status labels give complex work a calm hierarchy. Reusable controls and evidence patterns keep the product coherent across the journeys.
The opening view summarises useful preparation: permitted source evidence gathered, possible matches surfaced and drafts assembled. It names the inputs and their freshness. A prepared draft is not presented as an approved accounting conclusion or a completed engagement.
Each item states the evidence available, the checks passed and the gaps that remain. Missing, stale, conflicting and untested states are distinct. A confidence percentage cannot substitute for an explanation of what supports the decision.
The mandate states what can run in the background: monitor permitted sources, suggest matches, assemble evidence and draft requests or summaries. The owner can inspect activity, change the scope or pause pending work.
Sending a client request requires approval of the exact recipient, scope and wording. Professional treatment and current-packet sign-off remain human decisions. Autonomous ledger writes, filing, billing and payments are outside the initial concept.
Secondary counts show decisions, waiting evidence and review work across Maya’s ten authorised clients. Each count opens a filtered queue; each prepared item opens the relevant client, period, evidence and owner. Firm-wide planning needs a separate grant.
Next test: Can an accountant explain what Thread has done, identify the checks still required, approve or reject the proposed next step and change or pause delegation without assuming an unreviewed draft is complete?
Matching supplier and invoice references can justify investigation. They do not establish a duplicate. Source records, versions and the reason for the suggestion remain visible beside the proposed action.
A missing March statement becomes a request with a specific account, period, recipient and purpose. AI can draft it; Maya reviews the exact wording and scope before it is sent.
The client’s response enters a verification step. Maya checks the entity, account, period and coverage. A wrong or incomplete file remains unresolved rather than clearing the blocker.
In this read-only connector scenario, Maya makes the correction in the source system. Paper Thread then refreshes the source and reruns the affected checks, retaining the decision and its evidence.
Next test: Can a preparer distinguish a suggestion from a confirmed issue, reject incomplete evidence and explain the correction without silently clearing a blocker?
The reviewer approves a specific packet built from specific source versions. Later changes preserve the old history while requiring a new decision for the affected work.
Missing evidence and untested scope remain visible in the packet. The interface does not imply that a polished summary means every accounting or tax conclusion has been established.
Internal sign-off, client consent, export, submission, acceptance and payment are different events. Completion requires the appropriate receipt rather than an optimistic status change.
The decision, reviewer, source versions and evidence manifest can remain together in an exportable record. The product’s value should survive beyond its own screen.
Next test: Can reviewers spot a superseded approval and explain exactly which work is approved, what remains outside scope and what external outcome has actually been confirmed?
Paper Thread is an independent product concept. The value proposition and workflows are ready to examine; their benefit remains to be proven.
A research and requirements model, jobs to be done, user stories with acceptance criteria, proposed journeys and an editable product design system with connected screen states. The portfolio also includes a working browser prototype with editable fictional data and a local CSV import.
No accountant interviews, observed-practice study, usability evaluation, live integration benchmark or paid pilot has validated this solution. The people, clients and numbers used in the screens are fictional fixtures.
Begin with an accounts-and-tax preparation-to-review journey for a small set of consenting limited-company engagements. Test with sample or appropriately authorised data in shadow mode before relying on any result.
Measure time spent reconstructing context, missing-evidence discovery, false alarms, avoidable review returns, total effort and successful recovery. Compare against the existing workflow rather than counting generated summaries.
The interactive prototype is available above. Its source connections and AI are simulated; its editable inputs, calculations, CSV import and browser storage work.
Browse the product screensPaper Thread makes the product proposition tangible: an overview that leads to useful work, evidence that remains attached to decisions and assistance that operates within clear human authority.
The working prototype makes the journey testable. The next step is to evaluate it with practising accountants and use their feedback to refine the workflow.
