How we test whether AI can carry part of your workflow
Part of every review answers a technical question. Can AI take on a bounded piece of the work, on your data and inside your controls, well enough to justify a production pilot? We answer it by testing on your own closed cases, because a vendor demo on clean sample data cannot tell you how it will do on your drawings, your scans and your exceptions.
Vera AI guide
One Vertical Tier's AI guide. Ask me about the workflow review.
ReadyAsk me how the test works, or how your data is handled.
Getting ready on your device.
Vera is an AI. By default she picks one of the answers the OVT team wrote in advance, matched to your question in your browser. A model that writes a new answer works only from those written answers, also in your browser, and can make mistakes. Either way, what you type stays on this device. If you use the microphone, your browser's speech service turns your voice into text. Please do not share confidential information here.
Five questions, answered with evidence
Each answer in the findings report points to a logged piece of evidence: a document, a sample, a test result or a sign off.
- Where does each piece of information the decision needs actually live, and can we reach it through an approved, read only route?
- Is that data complete enough on a real sample of your cases, and can each past case be tied to its final outcome?
- Run on 30 to 50 of your closed cases, does the AI step meet the accuracy your workflow owner set before the test?
- Who approves what the AI drafts, and what happens when it is unsure or a case does not fit?
- Can every action be reconstructed later, within your security, privacy and customer contract rules?
Four stages, each with a gate
| When | Stage | What happens on the technical side |
|---|---|---|
| Days 1 to 2 | MobilizeG0, ready to begin | We agree the workflow boundary and the one AI step to test: extract, retrieve, classify, check or draft. Your IT owner and workflow owner sign a read only access boundary before any data moves. |
| Days 3 to 5 | DiscoverG1, current state validated | Working sessions with your IT owner and workflow owner. We walk three recent cases end to end and map every system and access route. Your workflow owner signs the accuracy thresholds before anything is tested. |
| Days 6 to 10 | EvaluateG2, findings baseline | We run the AI step on your closed cases, at least a quarter of them exceptions, and score each one against the outcome on record. Every failure is kept with its cause, so a bad scan is not confused with a model error. |
| Days 11 to 15 | DecideG3 and G4, acceptance | Twelve criteria scored with evidence. An executive readout with one consolidated correction cycle, and a decision your sponsor can act on the same week. |
Twelve criteria in four groups
Each criterion is scored with its rationale. A high score must point to logged evidence. Some gaps hold the decision whatever the total: no approved access route, no named human approver, no way to measure the baseline, or processing your contracts or regulators do not allow. A sample of 30 to 50 cases gives directional evidence, so results come with their confidence range and a pilot measures again on live volume.
Data and access
- Source of record
- Access path
- Structured data quality
- Document readiness
AI task fit
- Task definition and ground truth
- Measured performance on your sample
- Exceptions and failure handling
Control and evidence
- Human approval path
- Security, privacy and client restrictions
- Evidence and traceability
Operations
- Baseline measurability
- Integration path and ownership
One 60 minute session
- The systems in the path: ERP, CRM, document stores, email, portals, spreadsheets
- The access route for each one, and who approves it
- Data classes, and where processing is allowed to happen
- Any AI service you have already approved, and on what terms
- What logging exists today
One 90 minute session
- Three recent cases, on screen, start to finish
- The common exceptions and who resolves them
- Who approves what, and at what limits
- Where the final outcome of a closed case is kept
- A list of 60 to 80 closed cases; we choose 30 to 50 with you
Advance, resolve first, or stop
Advance to a production pilot
Nothing is blocked and every remaining gap has an owner. You get a pilot boundary that has passed our Pilot Readiness Gate: users, systems, approval matrix, thresholds, evidence record and rollback. The pilot is scoped and priced separately.
Resolve first
The AI step is close, but something upstream holds it back, usually data quality or access. You get each fix with an owner, effort and date, and a plan to test again on fresh cases before any pilot is proposed.
Stop
A fair test showed the AI step will not meet your thresholds, or your contracts or regulations rule out the processing it needs. You keep the baseline, the systems map and the findings, and you do not pay for a pilot that would have failed.
A findings report with the decision on page one
- The access boundary: every system, route, account and approver
- Evaluation evidence: thresholds, results case by case, every failure and its cause, confidence ranges
- The human approval path and failure handling for the AI step
- A scorecard of twelve criteria with rationale and evidence
- A risks and controls register with owners and dates
- A pilot boundary, a remediation plan, or a stop rationale
Read only, inside your controls
- Least privilege. We reach only what the scope requires, through named accounts, with role based access and MFA where supported.
- Read only during the review. Nothing is written to a production system.
- Your IT owner and workflow owner sign the access boundary before any data moves.
- Processing stays in your own environment, or one you approve. No public AI tools, and no model training on your data.
- We sign NDAs and follow your access logging and retention rules. Samples are returned or securely deleted within 30 days of final acceptance.
- Restricted material (export controlled drawings, privileged files, personal financial data) stays out unless your approved environment covers it.
Typical starting points
Manufacturing and distribution
- RFQ to approved quote
- Custom configuration to approved quote
- Quality exception to approved disposition
- Customer specification to approved quality packet
- Customer demand change to approved schedule
Legal, wealth and insurance
- New matter intake to complete attorney packet
- New client documents to adviser ready plan
- Claims submission to adjuster ready packet
What the review does not include
Production configuration, integration build or deployment. Legal opinions, certifications or audit opinions: references to SOC 2, NIST AI RMF, ISO/IEC 42001 or sector rules describe readiness and control mapping only.
The review is a fixed fee engagement of 2 to 4 weeks, scoped in our proposal. Most first reviews take 2 to 3.
Book a Workflow Fit Call
A free 30 minute conversation about one workflow: the problem, the systems involved, your constraints and who owns the decision. At the end we tell you plainly whether a paid review is the right next step.
Monday to Friday, 8 am to 6 pm Eastern. One Vertical Tier, Inc., Columbus, Ohio.
Worth bringing
- A short description of the problem and the result you want
- The systems and tools involved
- Known constraints, such as timeline, compliance and a rough budget band
- The key stakeholders, including who owns the decision