By the time a customer reaches your AI agent, something has usually gone wrong. A flight was cancelled. A charge that doesn't match. An order heading to the wrong door. Answering them is the easy part. Someone still has to do the work.

Today we're introducing Ada Computer, a new capability that gives your AI agent the ability to use your business systems and software to complete tasks for customers, not just answer their questions. If a person on your team can do the work using a computer, your AI agent should be able to do it too.

A shopper asks which EV charger will work for their car through a Quebec winter. The agent checks the catalogue, matches the connector to the vehicle, prices the option with the rebate applied, and confirms the shopper's request to add it to the cart. Three kinds of tools, one conversation. Recorded in the Ada dashboard with a demo store and demo values. Tool setup is shown separately from the conversation.

What we're releasing

Ada agents have always been able to call tools. If you've built with Ada, you know them as Actions. Now with Ada Computer your agent has access to a full set of flexible tools to reach your data and get tasks done quickly and effectively.

Code tools run authored, sandboxed Python inside the conversation. Refund amounts, entitlements, prorated fees and eligibility windows come from your rules running as code, so the number comes from your policy rather than from the model. Your team can write and test a code tool in the dashboard, or describe it to a coding agent and let it draft the Python.

MCP tools connect a server your company already has, through the open Model Context Protocol. Point Ada at the server by URL and every tool it exposes is discovered. Each one starts off until you turn it on, giving you full governance.

API tools connect the APIs your business already runs, so the agent starts from the facts: the real order, the real balance, the real booking.

Why we built it

When we looked at what separates a good conversation from a resolved one, the answer was rarely the conversation. It was everything behind it. Language models are very good at understanding a customer and deciding what to do next, and we kept seeing the same three things get in the way.

Models were doing arithmetic. They were asked to parse dates, match order numbers and compute refunds, and every time they did, accuracy dropped and latency climbed.

Every action was a project. What an agent could do was capped by the integrations a team had time to build, so resolution stopped where the backlog started.

Teams wanted more control, not less. When the agent acts, which systems it may touch, and a record of every step it took.

Ada Computer is our answer to all three. The Ada Reasoning Engine™ decides what to do next. Your APIs, your code and your MCP servers carry it out. Then the engine takes it from there.

One task, start to finish

Consider a cancelled flight and three travellers who need to get to New York tonight.

The Reasoning Engine understands the request and plans the path: verify, find options, check entitlement, confirm, rebook. An API tool reads the reservation: passengers, bags, fare class. A code tool applies the airline's entitlement rules and returns exactly what these travellers are owed. The AI agent confirms with the customer at a step you defined.

"The 6:10pm via Montreal works for all three. Shall I book it?"

On a yes, an MCP tool rebooks through the carrier's ticketing system and reports the result back. The agent closes the task. Every acting step hands back to the engine, so the loop is always the same: decide, act, check, decide.

Illustrative scenario with sample data. Entitlements vary by airline and jurisdiction.

Control stays with your team

Giving an agent the ability to act raises the bar on governance, so we designed for it from the start.

You define when it acts. Playbooks place each tool call at a specific step, with the checks you want in front of it. Add an Ask step before anything that writes.

You keep judgment out of policy. Amounts, entitlements and windows run as authored code, in a sandbox, with no network access unless you allow it.

You see what ran. Every tool run is traced in the conversation view, in sequence with what the agent said and did around it.

MCP tools start off and are enabled one at a time. Your systems' own permissions still apply to every call the agent makes.

Where this goes

AI customer service started by answering questions. Now AI agents can do the work.

Ada Computer is the action layer of the AI agent, and it will keep growing. Our direction is an agent that can carry your most complex customer journeys from start to finish, across more systems, over longer stretches of work. Every task it completes also feeds the loop that makes it better: what ran, what worked, and what to coach next.

Start with one task

Pick one thing your customers ask for often that your agent explains today but doesn't finish. Note the systems it touches and the rules it follows. Then bring it to us. We'll show you what your agent needs to complete it end to end.

Want more info? Visit the Ada Computer page.

Read the documentation