Assistants and agents

For work that has to read, decide and act: an assistant that answers from your documents, or an agent that carries a task across your systems, built so you can check it.

A demo is easy. What is hard is the tenth week: the answer that sounds right and is not, the task that loops, the bill nobody was watching. Most of the work in a real build goes into the parts a demo skips.

Describe the task in writing

 ┌─────┐   ┌──────────┐
 │ ◉ ◉ │──▶│ ━━━━━━━  │
 │  ─  │   │ ━━━━━    │
 └─────┘   │ ━━━━━━   │
           └──[ ? ]───┘

Example workflow: an assistant that answers from your documents

Runs on its own 1. A question or request arrives From a client, a colleague or another system.
Runs on its own 2. It reads your documents Only the ones this request needs.
Runs on its own 3. It drafts the answer or action Built to draw on those documents, not the model's memory.
Waits for a person 4. You approve what matters Actions that commit you wait for a person, as agreed in the quote.
Runs on its own 5. It acts and keeps a record What it did, and from which documents.

Alert path, from steps 3 and 5

Alerts someone A check fails or the spending cap is reached. It stops and a person is told.

Runs on its own
Waits for a person
Alerts someone

What is included

The simplest design that does the job: often an assistant, sometimes an agent
A written test set built from your real cases, rerun when the build changes
A spending cap and a stop condition on each run, set to cut off a run that loops or overspends
A human approval step on the actions we agree need one
A log of what it did and what it read to do it

How it goes

We write down what it must do, what it must not do, and how we will know it works.
We build a pilot on your real documents and run it against the test set.
It starts with little autonomy. More is added in steps, after the tests pass.
It goes live with alerts in place. A retainer covers re-testing when the model changes.

Price

A fixed price for the scope in the quote, in writing before any work starts. The pilot is the first stage: you see it run before you commit to the rest. Usage costs from your AI provider are billed to you by the provider and estimated in the quote; the estimate is not a cap on what the provider charges.

Will it make things up?

We build it to answer from your documents and to mark what it cannot find instead of filling the gap. That reduces invented content; it does not eliminate it, which is why the approval step and the tests stay.

Assistant or agent?

An assistant answers and drafts, and a person acts. An agent takes the actions itself. Agents cost more to build and to run, so we use one only when the task needs it.

What does it cost to run?

It depends on the volume and on the model. We measure cost per step during the pilot and put a written estimate in the quote for the rest of the build. It is an estimate: your provider bills actual usage.

Describe the task

What happens next

A person reads it.
You get a written first read, usually within one business day.
If it is worth doing, a written quote follows.

Prefer email? Write to info@acceleratech.ca.

Other services

Product names are trademarks of their owners. Naming a tool means we use it, not that its vendor partners with or endorses us.