New kinds of input
A new form layout or email style can break extraction that worked at launch.
AI operations
AI that people check, after it goes live.
Most AI projects stop at launch. We review outputs, resolve exceptions and improve the workflow for as long as it runs. Reviews happen on a schedule agreed with you.
An example review pathAI + human review
Flagged output→Review→Resolve or escalate
A
Agree the review rules.
Specify the criteria that send an output to a person.
B
Record the decision.
Capture the resolution and route anything outside the agreed scope.
C
Feed findings into the build.
Use repeated issues to guide improvements.
Why it matters
AI workflows drift
without review.
Inputs change, new edge cases appear and quality slips quietly. Human review catches it before customers do.
Silent errors
Without a review queue, a wrong output looks the same as a right one.
No one owns it
After handover, small issues pile up because nobody is responsible for the workflow.
What we do
Review, resolve,
improve.
Review outputs
Check work against the standards that matter to the workflow.
- Defined quality and accuracy checks
- Review of uncertain or incomplete outputs
- Escalation to an agreed decision-maker
You receive: reviewed outputs and a record of each decision.
Resolve exceptions
Handle the cases that need more than an automated next step.
- Failed-run triage and follow-up
- Incomplete records and ambiguous data
- Documented resolution and escalation paths
You receive: exceptions resolved or escalated to the right owner.
Improve performance
Use recurring issues to inform the next implementation change.
- Issue patterns and review findings
- Updates to prompts, workflows and evaluations
- Documentation of changes and unresolved issues
You receive: regular findings and the changes made from them.
Setting up support
Coverage built
around the workload.
Start with a workflow review. Agree the queue, review criteria, expected volume, service hours, access and escalation path.
- What is reviewed and what is excluded
- Who resolves each type of exception
- How work and decisions are recorded
- When coverage is available
- How often findings are reviewed with you
A review cycle
What happens
each cycle.
Each review cycle follows the same four steps.
Collect
Gather flagged outputs and a sample of routine ones.
- Items flagged as uncertain
- A random sample of completed items
- Failed or incomplete runs
Check and resolve
Review each item against the agreed criteria.
- Approve, correct or escalate
- Record the reason for each change
- Escalate out-of-scope items to your owner
Find patterns
Look for issues that repeat.
- New input formats
- Recurring corrections
- Rules that need updating
Improve and report
Change the workflow and tell you what changed.
- Prompt, rule and test updates
- Findings shared with you
- Open issues and next actions
You receive: a short findings report each cycle.
Who it’s for
When to add
ongoing support.
You have no AI team
Nobody in-house has time to watch the workflow every week.
Errors are costly
A wrong output reaches a customer, a record or an invoice.
Volume is growing
More items means more edge cases and more to review.
Questions
AI operations,
answered.
What is human review of AI outputs?
A person checks the outputs the workflow flags as uncertain, or a sample of all outputs, against agreed criteria, and either approves, corrects or escalates them.
Can you support a workflow you did not build?
Yes. We start with a workflow review to understand how it runs, then agree the review rules and access.
How do we know quality is holding up?
Findings are recorded and shared with you, so you can see what was flagged, what was changed and what remains open.
Who has the final say on an exception?
An owner on your side, agreed in advance. Formals resolves what is in scope and escalates the rest.
How is support priced?
By expected volume, coverage hours and review depth, set out in a written proposal.
Work with FormalsNeed support
Need support
after deployment?
Tell us what is running and where human support is needed.
Discuss AI operations