How it works
From the AI work you can't see to agents your company runs on.
An engagement with Sonsara has four steps. We find the AI work your people already do, build the best of it into tested company agents, put them where your team works, and then keep them running and improving.
The engagement
Discover. Build. Roll out. Run.
Each step hands something finished to the next, and the last one doesn't end. Agents are looked after for as long as they run.
Find the AI work already happening
We read the logs your tools already keep to see who uses AI, for what, and on whose accounts. Then short interviews with the people behind the work worth keeping.
Turn it into company agents
Each agent is built from a written spec, tested on real cases, and approved by its owner after a test run. Then it goes into your company vault.
Give your team access
People reach their department's agent in Slack, Teams, Claude or ChatGPT. Same agent, same controls, wherever they ask.
We keep them live and improving
We watch every agent, fix what breaks, keep connectors working, tune cost and model choice, and improve answers as real use comes in.
01Discover
See how your people really use AI.
Some people have built powerful agents nobody knows about. Some use personal accounts and tools nobody approved. Some barely use AI at all. Discovery shows you all three.
- Read-only access to the audit logs your identity provider and work tools already keep. Nothing is installed on anyone's laptop.
- We read metadata only: which system, which AI tool, when and how often. Never prompts or content, and people are pseudonymised.
- First findings in days: which AI tools are in use, by which teams, for which jobs, and what's risky.
- For the work worth keeping, a short voice or chat interview with the person who does it. The result is a written spec.
02Build
Built, tested, approved, then into the vault.
We rebuild each piece of work as a company agent: on company credentials, with guardrails, and tested before it ever runs on its own.
- Each agent follows a written spec: what starts it, the steps, the tools, the data and who owns it.
- Dry runs and evals on real cases before go-live. Failing evals block approval.
- The owner watches a test run and approves it. Only then does the agent go into the vault and run.
- Connections belong to the company, not a person, and each agent gets only the access its job needs.
Supplier reconciliation
finance agent · weekdays 09:00 · v1
- Interview with the person who did it
- Evals on real invoices
- Company credentials, scoped
- Every run logged
- Test run approved by the owner
03Roll out
Your whole team gets the agents your best people built.
Agents live where people already work. Each department has one agent to talk to, and it hands the job to the right sub-agent behind the scenes.
- Slack, Microsoft Teams, Claude and ChatGPT. The same agent, under the same controls, on every one.
- People ask their department's agent, for example "@Finance agent", and it runs the right job.
- Access is set per agent: owners, people who can change it, and people who can use it. Scheduled agents also run on their own.
- Light users get the same shortcuts as power users, without having to build anything.
@Finance agent can you run this week's supplier reconciliation?
Done. 46 supplier statements matched to Xero. Two need you:
- Acme Logistics: invoice missing from Xero
- Brightline: paid twice in March
04Run
Run by us, with a 30-minute fix.
We're the MSP for your agents: like the IT firm that keeps your laptops and email working, we don't just build agents, we keep them running and take the call when something breaks. Tools change, connections expire and the work moves on, so we look after every agent in the vault for as long as it runs.
- Every run is watched. If an agent fails in production, it is ours to fix, within 30 minutes.
- Connectors kept working. When one disconnects, the agent is marked down and its owners are asked to reconnect, right in Slack or Teams.
- Cost and models tuned over time, so each job runs on the model that does it well for the least spend.
- Evals and answers improved as real use comes in, and skills refreshed from how people actually use them.
- Every change is a new version that is test-run and approved before it replaces the old one.
FixedRenewal radar: Gong disconnected. Owner asked in Slack, reconnected, missed run re-done.
TunedCall follow-ups moved to a smaller model after evals matched. Same answers, lower cost.
ImprovedSupplier reconciliation v2 approved: handles credit notes, two new eval cases.
Who does what
What you do, and what we do.
Your part is mostly decisions and a little of your people's time. The build, the testing and the upkeep are ours.
01Discover
Grant read-only access to your identity provider and work tools. Tell your people what's happening. Each person whose work we keep gives a short interview.
Read the logs, map AI use by team, flag the risks, run the interviews and write a spec for every agent worth building.
02Build
Choose which agents to build. Each agent's owner watches a test run and approves it.
Build each agent from its spec, connect it on company credentials, write the evals, test it and put it in the vault.
03Roll out
Add the agents to Slack, Teams, Claude or ChatGPT, and decide who can use, change and own each one.
Set up each department's agent on every surface you use, under the same controls.
04Run
Reconnect a tool when asked. Approve new versions. Tell us when the work changes.
Watch every run, recover failures, keep connectors working, tune cost and models, improve evals and refresh skills from real use.
The first weeks
What the start of an engagement looks like.
The pace depends on your systems and your people's diaries. The order doesn't change.
First days
Connect and see.
A short kick-off, then read-only access to your logs. First findings arrive in days: the AI tools in use, the teams using them and the risks.
Next
Interview and build.
You pick the work worth keeping. The people behind it give short interviews, and the first agents are built, tested and approved by their owners.
Then
Roll out and keep going.
Agents go live in the tools your team uses. From there we run them, and each new round of discovery adds the next set.
FAQ
Questions about an engagement.
How much of my team's time does it take?+
Very little. Discovery runs on the logs your tools already keep. The only time we ask of people is a short voice or chat interview for each piece of work worth turning into an agent, and an owner's approval of a test run.
Do we need to install anything?+
No. Discovery uses read-only access to your identity provider and work tools, with one app consent or admin key per system. Nothing goes on anyone's laptop.
What happens when an agent breaks?+
We're watching every run. A production failure is ours to fix, within 30 minutes. If the cause is a disconnected tool, the agent's owners are asked to reconnect it in Slack or Teams.
Who owns the agents?+
Your company does. Agents run on company accounts and company-held connections, with named owners. When someone leaves, their agents stay.
Keep reading
Each step in more detail.
Book a demo
See the AI work already happening in your company.
Read-only access, nothing to install, first findings in days.