Operator agents
Uses your software step by step, the way an operator would.
What's included
The agent works through your product's own screens and APIs, the same way a trained operator would. Each workflow is scoped, built and accepted on its own.
Ship AI in your product.
An agent that operates your software; a person confirms what it isn't sure of.
- Customers ask for AI in the product; the roadmap has no room to build it.
- The demo impressed everyone; the accountants asked what happens when it's wrong.
- Every customer's data looks different; a one-size model doesn't fit.
10B+messages · founder's prior work
6warehouse domains · founder's prior work
1embedded agent · in progress
What we build into your product
Verification queue
Uncertain steps go to a person; the answer is reused.
What's included
Confidence thresholds per judgement; a queue your customer's staff work through; each answer applied to the current batch and offered as a rule.
Rules per customer
Confirmations become rules, kept per customer.
What's included
Closed rule types with executable handlers; adding a rule is data, not code. Each of your customers keeps its own rules, reviewed on a cadence they set.
Evidence trail
Every step recorded, auditable by your customer.
What's included
Run, step and decision records; what was read, what was proposed, who confirmed, what was written. Exportable for audit.
Deployment package
Runs in your environment. We host nothing.
What's included
Container images, migrations and configuration for your infrastructure; model access through your approved provider; nothing sent to us.
Evaluation and acceptance
Real cases with right answers; acceptance on measurable criteria.
What's included
An evaluation set built with your team; automated runs on every change; acceptance on pass rate, error bounds and throughput agreed before the build starts.
How it works
- Your product's dataFacts from its own records
- Agent proposesA judgement, with evidence
- Person confirmsVerification queue
- Execute and recordWritten to your product, logged
- Rules accumulatePer customer, reviewed on a cadence
Step 05 feeds back into step 01 · the next file from the same source is handled automatically
Where it fits
- Accounting and tax software
- ERP and operations software
- Document-heavy vertical SaaS
- Internal tools
Your customers' data stays theirs
- Your environment
- Per-customer isolation
- Your model channel
- In the contract
Built into your product. Checkable by your customers. Owned by you.
How this compares with building in-house, generic AI features and AI consultancies
| Build in-house | Generic AI features | AI consultancies | Unvessel | |
|---|---|---|---|---|
| Knows your product | Yes | No | Learns on the job | Built inside it |
| When it's unsure | Depends | Guesses | Depends | Asks a person; remembers the answer |
| Per-customer behaviour | If you build it | One size | Depends | Rules per customer |
| Verifiability | If you build it | Black box | Depends | Every step recorded |
| Ownership | Yours | The vendor's | Usually yours | Yours; runs in your environment |
| Way in | Hiring | Low | High | A two-to-three-week scoping |
How to start
- ScopingTwo to three weeks: one workflow, its acceptance criteria, a fixed price
- Build, phase oneOne workflow live, with the queue, rules and records
- Managed operationsFrom go-live, monthly; further workflows by phase
- We don't sell you a chatbot.
- We don't train models on your customers' data.
- We don't host your product's AI; it runs where your product runs.
Questions
Which model?
The provider and region you approve. The agent is built so the model can be swapped; the queue, rules and records don't depend on it.
What happens when the agent is wrong?
Uncertain judgements never execute; they wait for a person. Wrong confirmations can be withdrawn, and every step is on record, so the effect can be traced and reversed.
Does it learn across our customers?
No. Rules are kept per customer. Cross-customer generalisation is a separate decision, made by you.
How is it priced?
Scoping at a fixed fee; each workflow phase at a fixed price with acceptance criteria agreed first; managed operations monthly.
See the embedded AI case in progress.
Start with a scoping
Tell us what your product does and which workflow your users would hand to an agent first. We'll reply with how we'd scope it — or tell you honestly if we're not the right firm.
info@unvessel.ai