How a global services leader cut the time to build an AI agent from 10–15 days to hours

The Challenge

Prove it inside first, then take it to clients

The firm is a global business-services leader with more than 100,000 employees and over $5B in annual revenue. Its IT organisation supports that workforce across service desk, end-user computing, identity, collaboration and infrastructure teams.

Leadership has set a clear direction: autonomous IT by 2027. The plan was to prove AI agents inside its own IT teams first, as "client zero", and then take what worked to its clients. An internal AI team had already built around 30 agents, but each one took 10–15 days to build. Knowing what to build next meant mapping value streams by hand, which typically took 3–6 months of interviews before a single opportunity surfaced.

The question was not whether agents could work. It was which ones to build, and how to build them fast enough to keep up with the roadmap.

The Deployment

Sixteen teams observed, from kickoff to live agents in under eight weeks

Worktrace was rolled out silently across 16 IT teams and more than 100 users, with no integrations into business systems. The desktop app captured real work as it happened, with personal data redacted on the device. Within three weeks of recording, the platform had mapped the service desk's work to the step level and produced its first build-ready agent blueprints. Agents were running on live data in under eight weeks from kickoff.

What Worktrace Revealed

Four findings that changed what got built, and how fast

1. A prioritised roadmap of automations, sized by real hours

7,576 hrs: estimated capacity a year across six automations

Of the 8,200+ hours observed, 88% sat in work an agent could take on. Worktrace ranked that work into six automations, each tied to a clear problem and an estimated annual saving.

• 3,502 h: Queue and staffing monitoring. Scheduled agents watch queue dashboards and staffing all shift, and alert only on a breach.
• 2,468 h: Ticket task creation and tracker updates. Tasks are created from templates and mirrored to trackers, ending the re-keying.
• ~640 h: App packaging and testing. An agent builds, tests and stages each package, and the engineer approves the release.
• 427 h: Bulk device deployments. Device lists in spreadsheets become validated groups and assignments, with an evidence trail.
• 293 h: Email and chat triage. Threads are summarised, ticket IDs extracted and replies drafted.
• 246 h: Incident closure. Closure notes and evidence are pre-filled for one-click approval.

Blueprint estimates from observed sessions. Packaging uses the midpoint of 45–62 hours a month.

2. A fact base that no runbook had captured

28–30%: of desk effort found, against a 10% target

A Mission set to give the service desk back 10% of its manual effort found an estimated 28–30%, about three times the target. The data also showed where the time was really going.

Packaging an app took 24 minutes of hands-on work but nearly 16 hours from start to finish, so waiting outweighed working. Every identity or MFA reset meant checking four systems by hand before anything changed. Installs pushed without a status check came back as new tickets.

The data was just as clear about what not to automate. About a third of the desk's time went on calls, huddles and hands-on sessions, which stayed with people. Some proposed agents would have duplicated reminders the ticketing tool already sent, and the fix there was a change in cadence rather than a new agent.

3. Missions turned leadership goals into better service, not just saved hours

12: Missions tied to leadership goals

Each Mission started from a plain-language goal, such as fulfilling 90% or more of software install requests with zero human touch. Worktrace answered with ranked agents grounded in that team's recorded work.

The biggest unlocks went beyond productivity. Every closed ticket can now be quality-audited instead of a sample. Tickets sitting idle past their follow-up window are flagged every 30 minutes. Verified installs mean fewer repeat tickets. Every app package is tested before it ships, and each change carries one evidence pack for audit.

4. Every review made every agent better, and the next one faster

1 wk → 3–4 h: agent build time, from first agent to today

The firm's engineers reviewed each agent and flagged what didn't fit. Instead of patching agents one at a time, Worktrace folded that feedback into a shared org context and regenerated every blueprint from it.

That context records which tools each agent can read, write or send to, and what is off limits. When an integration isn't available, agents keep working and list the access that would unlock more. Rules moved out of the model and into editable config, which made agents predictable enough to run. New agents inherit earlier lessons by default.

The result compounded quickly. The first agent took about a week, the second a day, and revisions now land in 3–4 hours.

Value Realised

Three unlocks for the firm

Discovery in weeks, with no manual mapping

3–6 mo → <3 wks: time to a mapped value stream

Value streams and 32 build-ready blueprints came from the first three weeks of observed work, replacing months of interviews and shadowing.

Agents built at the pace of the roadmap

10–15 days → hours: time to stand up an agent

Against a 10–15 day internal baseline, agents now go from blueprint to running in hours, keeping pace with the 2027 autonomous IT goal.

A tested proof point to take to clients

3 agents: running live, operated by the firm's own engineers

Three agents were handed over and run live by the firm's engineers. That gives the firm a client-zero story it can take to its own customers.

The Bigger Picture

Next unlock: L2 and L3 work, where APIs stop

Work like app packaging needs a virtual machine and a console, not just an API. The next step pairs Worktrace workflows with computer-use agents that will run in their own VM, using the same tools an engineer uses. A richer capture pipeline now records the reasoning behind each step, not just the clicks.

From here, the firm plans to scale to 200 users and take proven agents to its own clients, with a roadmap built from evidence rather than assumptions.