← All briefings

AI Implementation & Workforce

Your AI Rollout Needs an Implementation Bench, Not More Licenses

The pilot works while its builder is in the room. What happens when they leave? Build a small implementation bench that connects company knowledge, employee agents, and dependable daily work.

The pilot works. Its builder is on vacation.

Imagine an operations team that has built an agent to prepare internal request briefings. On Friday, the demonstration looks excellent. By Tuesday, an intake field has changed, a policy has been replaced, and a colleague cannot explain why one request was routed differently. The person who built the workflow is away. Everyone returns to the old spreadsheet.

This is an illustrative scenario, not a Delphi client case. The failure is not necessarily the model. The company has created a useful capability without creating the people and operating knowledge needed to sustain it.

Before buying more access, ask who can carry an agent from a convincing demonstration to dependable daily work. That is the purpose of an implementation bench: a small group with the business understanding, technical ability, review authority, and protected time to make AI useful after launch day.

The source signal: training is moving closer to delivery

Anthropic's Claude Frontier Academy announcement commits $100 million and sets a goal of training 10,000 Frontier Deployed Engineers by the end of 2027. Its described program combines a simulated enterprise deployment, a graded practical assessment, and a 12-week residency leading a real project inside the participant's organization, followed by another assessment.

Those are announced commitments and program design, not evidence that 10,000 engineers have already graduated or that the investment has produced a measured business return. The accessible announcement does not establish its publication date, so it should not be treated as a confirmed launch this week.

The useful management signal is the training model. Practice includes use-case selection, security review, and handover, not just learning the interface. A smaller business does not need to reproduce a vendor academy. It can adopt the same discipline: give capable people a real workflow, assess their work, and make operational readiness part of completion.

Build three responsibilities, not another innovation silo

Start with three responsibilities: a business owner who defines the outcome, an implementation lead who builds and maintains the workflow, and a reviewer who determines whether the result is fit to use. In a small company, one person may hold more than one responsibility, but consequential decisions may need an independent reviewer. Combining titles should not remove the control.

For the request-briefing agent, the operations manager owns the process. The implementation lead connects approved intake records and maintains the workflow. A team supervisor reviews the prepared briefings and escalations. Security, privacy, legal, and system owners join when the proposed access or use requires their expertise.

The bench should sit close to the department's work. Do not create a separate team measured on demonstrations while operations inherits the maintenance. Give the implementation lead protected delivery time and name a backup. Otherwise, the pilot becomes another responsibility carried by the employee who was already too helpful to say no.

Make the Company Brain part of the assignment

The first deliverable is a source map, not a longer prompt. For internal request triage, identify the current intake definitions, routing rules, service commitments, escalation policy, and team ownership records. Give each source an owner, an effective date, an access boundary, and a rule for resolving conflicts.

That is the Company Brain at workflow scale: approved operating context an employee agent can use without having to rediscover how the company works. It can begin with a small set of maintained records. It does not require connecting every document or exposing every employee's information.

Keep knowledge separate from enforcement. The source map explains the approved process. Tool permissions determine which records the agent can read or change. In the first version, let it classify requests and prepare source-referenced briefings in a restricted review area. Disable customer contact, external publishing, financial changes, and updates to systems of record.

Assess practical judgment, not prompt confidence

OpenAI's November 5 account of Australian university collaborations describes UNSW's Can, Can't, Must assessment framework: educators specify when AI use is permitted, prohibited, or required. These are education examples reported by a vendor, not proof of workforce performance. But the distinction is useful when designing employee training.

Adapt it to the actual role. An employee can use the agent to prepare a briefing from approved records. They cannot treat an unsupported recommendation as policy or let the agent grant itself new access. They must flag unresolved source conflicts and obtain the designated review before a consequential handoff.

Then assess those behaviors with protected test cases. Include an ordinary request, an outdated policy, a missing owner, restricted information, and an instruction inside an intake document that asks the agent to bypass review. Specify the expected behavior before testing. The right answer may be a clear escalation rather than a finished briefing.

The handover is the real graduation

A workflow is not ready because its author can make it work. Before expansion, ask the backup operator to run it, diagnose a failed input, pause it, and restore the last approved configuration without relying on the builder's private chat history.

This is Delphi's recommended readiness test, not a vendor certification. Keep the handover record somewhere authorized colleagues can find and maintain it. Sensitive configuration and credentials belong in approved secure systems, not in the operating instructions.

  • Mandate: purpose, owner, reviewer, backup, allowed actions, and disabled actions
  • Context: authoritative source locations, access rules, freshness checks, and conflict handling
  • Release: approved workflow version, representative tests, known limits, and review date
  • Operations: start, pause, failure diagnosis, escalation, and recovery instructions
  • Control: permission inventory, approval boundaries, logs, and tested access revocation
  • Economics: baseline, usage costs, support effort, review time, and accepted-result measures

A first-month playbook for the implementation bench

Choose one recurring internal workflow whose output can be reviewed before it affects a customer or a consequential decision. Set the pilot's decision date and acceptance standard before building. The first month should produce a maintainable operating pattern, not a promise to automate a department.

  • Week 1: name the three responsibilities and backup; observe the current process and record handling time, review effort, rework, and quality
  • Week 2: establish the Company Brain source map, narrow the agent's mandate, and configure read-only or draft-only access
  • Week 3: run ordinary and exception cases; correct source gaps and verify that restricted actions remain blocked
  • Week 4: run a limited pilot alongside the existing process, then have the backup complete the handover test
  • At review: compare accepted briefings and total handling effort, including supervision, rework, software, implementation, and support costs
  • Before expansion: decide whether to continue, revise, or stop; assign source maintenance and explain how any released capacity will be used

Buy capacity you can actually sustain

More licenses may be appropriate once a workflow is useful and supported. They do not resolve an absent owner, stale knowledge, an untested approval boundary, or dependence on one builder. The investment decision should include the people and maintenance needed to keep the capability working.

Every company needs a Company Brain, and every employee needs an AI Agent connected to the context of their role. An implementation bench makes that ambition practical by turning local experiments into maintained, reviewable work.

At your next AI review, ask the question the demonstration rarely answers: if the builder were unavailable tomorrow, could an authorized colleague keep this workflow useful, controlled, and accountable? If not, build the bench before expanding the rollout.

Your next decision

Turn the briefing into an operating advantage.

Plan Your AI Implementation Bench