Customertimes turned Claude Code into a delivery service it sells

Customer Overview

Governed, in-perimeter AI operations for a technology consultancy working in regulated industries: Claude deployed as the complete reasoning layer for software delivery, QA, knowledge work and AI cost governance, running under corporate SSO inside the company perimeter.

This is a first-party deployment: Customertimes is both the Anthropic partner and the deploying organisation. The platform runs across our own delivery business and is sold to our clients on the same build, which makes the internal deployment the reference implementation rather than a demo.

Industry

Professional services

Published

15 September 2026

All figures come from platform telemetry, the release record and the delivery organisation, and can be evidenced on request.

• Claude Code in customer delivery current rate across live engagements

  • Live on 12 customer delivery engagements, each running its own Digital Developer instance scoped to that client's Jira, repositories and credentials. Claude Code is currently running at around 58 issues resolved and 12 pull requests opened a month across them.
  • Per engagement, Claude Code works from Jira and GitHub under its own identity: analysis with clarifying questions, extended thinking on diagnosis and design, implementation on a feature branch, a pull request in every repository it changed, and a dated activity log of what it did. For Salesforce work it deploys to a sandbox, runs Apex tests and a bounded fix loop before opening the pull request.
  • In a blind pilot on a live customer project, Claude Code matched the human developers' diagnosis from the code alone.
  • Deployment is credential-scoped per engagement; production writes are refused unless explicitly enabled.

  Not every issue is a code change, which is the point. In a typical month the work divides into roughly 12 implementation runs, 19 planning runs and 85 replies. Claude Code is asked to review an architecture, produce an implementation plan, diagnose a stuck integration or answer a question in the tracker as often as it is asked to ship a change — and it opens a pull request on every repository it does change.

• Claude in QA automation

  • Claude drives web and mobile UI tests from natural-language test cases, with no scripts, over Playwright for web and Appium for mobile — self-healing locators, crash detection, parallel execution, results pushed back to the test-management system.
  • Running at around 110 quality runs and 28 test suites generated a month, with customer QA and project management having formally approved the tool on one account.
  • Recorded tests replay step by step with no model in the loop, so the model is spent on authoring and repair rather than repetition.

• Consumption model

  • Zero consumer AI seat subscriptions. One hundred per cent of Claude usage is metered API consumption on AWS Bedrock and the Anthropic API — the platform replaced per-seat tooling outright.
  • Consumption scales with delivered work rather than headcount: growth in Customertimes' delivery business is growth in Claude consumption.
  • Requests are routed to the cheapest capable tier — Haiku 4.5 on always-on paths, Sonnet for the bulk of production traffic, Opus reserved for work that needs it — with spend metered per user and per workspace and visible to leadership in the admin console.

• Scale and adoption platform telemetry, August 2026

  • 235 monthly active people in August 2026, up 61% on July, with 100 of them first-timers. 7,273 human messages, up 151%. 2,028 chats, up 141%. 1,319 automations firing, up 231%. Company-wide enablement completed in late August 2026.
  • In the same month the platform produced 171 tasks, 584 files saved into workspaces with 278 revisions on top, 94 presentations and decks, and 280 insights reports.
  • Four Claude models carry production traffic across three tiers. By share of model spend: Sonnet 5 87%, Sonnet 4.6 9%, Opus 5 3%, Haiku 4.5 under 1% — the routing policy visible in the ledger rather than asserted. The platform processes around 4.8 billion tokens a month, with 70% of reads served from cache. By share of spend the largest workloads are the conversational agent 62%, the code interpreter sub-agent 14%, and quality runs 10%. One hundred per cent of platform reasoning runs on Claude; no other model family is used for reasoning anywhere in the platform.
  • Ten first-party MCP servers, 13 integrated corporate systems, 220+ tools.

• The integration surface, in use

  • The platform serves around 27,800 data and document records a month across integrated corporate systems, Salesforce the largest single source at some 14,600 of them.

  Monthly tool calls, for every source with traffic that month. One call returns many records, so these counts are smaller than the figure above. Knowledge search and workspace files are internal platform surfaces rather than corporate systems:

  • Jira1,295
  • Salesforce754
  • Knowledge search internal585
  • SharePoint519
  • Outlook449
  • Workspace files internal428
  • Teams238
  • TestRail161
  • Confluence160
  • GitLab97
  • Azure Boards73
  • GitHub55
  • Azure Test Plans51
  • Xray34

• Governance

  • SSO-only authentication; workspace membership is the access boundary; tools execute under the requesting user's own connector grants; a departing employee's indexed content is purged.
  • Every Claude call is on the record: MCP traffic passes through one gateway that logs it, agent runs keep a dated activity log of what the agent did and which tools it used, requests are metered per user and per workspace, and administrative actions are written to a tamper-evident HMAC hash chain.
  • An approval gate on every agent write action, a two-tier prompt-injection filter on every call, and Haiku 4.5 running as an always-on safety judge gating every write to per-user memory.
  • No corporate data reaches public AI services; outbound web queries are screened for corporate content before they leave.

• Business outcome

  • Where Customertimes sells a unit of QA or maintenance capacity and delivers it with a smaller human team plus Claude, the difference is Customertimes' own margin. The client receives what they contracted for; the productivity gain is captured by the company that built the platform.
  • The same platform is sold onward: the internal deployment is the reference implementation for an AI-native managed-services offering now in market.