The person everyone waits for
Consider a familiar scene. A proposal is nearly finished, but nobody will send it until the operations manager looks at it. She knows which pricing sheet is current, which delivery promise needs a second approval, and why an apparently standard customer request is not standard at all. The template is in the shared drive. The judgment is still in her head.
This is an illustrative scenario, not a Delphi client case. It captures a useful starting point for employee agents: the recurring work that depends on one experienced person translating company knowledge into a decision.
The goal is not to copy that person's judgment into a prompt and remove them. It is to make their repeatable know-how available to the team while preserving their authority over exceptions. That requires something more precise than a document library: a tested AI skill connected to the Company Brain.
The source signal: knowledge becomes reusable instruction
In its September 16 announcement, Google described Workspace skills as reusable prompts that guide Gemini with a team's rules, templates, and reference files. Teams can collaborate on skills in Docs and test them in Workspace Studio. The examples include consistent proposals, brand standards, status updates, and invoice review.
The availability details matter. That announcement limits access to organizations whose administrators have enabled the Gemini Beta setting. It describes Marketplace publishing and administrator curation and distribution as coming soon. Leaders should confirm their own tenant's availability rather than treat the announcement as a universal rollout.
The practical implication is broader than one vendor. Company knowledge can become reusable instruction inside work, rather than material an employee must rediscover and explain in every chat. But reusable does not automatically mean reliable. An unclear rule becomes an unclear rule used repeatedly.
Separate the Brain, the skill, and the authority
The Company Brain holds approved operating context: current policies, pricing, templates, process rules, and decisions, with owners and access boundaries. A skill describes how to use a selected part of that context for one recurring task. The employee agent applies the skill within a defined role. The permission system determines what it may actually do.
These are different responsibilities. A proposal-preparation skill can explain how to check an offer against approved terms. It should not grant the agent authority to approve a discount or send the proposal. Written instructions help guide behavior, but access controls and approval mechanisms must enforce the boundary.
Keeping those layers separate also makes maintenance possible. A pricing change should update the authoritative source. A new review requirement should update the workflow and its tests. Neither should depend on finding every employee's private prompt history.
Give the first skill a small, useful job
Anthropic's October 1 Barclays case study offers a concrete design reference. It describes a retrieval-augmented Colleague Knowledge Assistant that helps employees find information, and a separate Global Markets platform that classifies, enriches, and routes incoming emails. These are bounded forms of assistance, not evidence that an agent autonomously runs the bank.
The announcement reports more than 16,000 colleagues adopting the knowledge assistant and over one million searches. Those are vendor-reported activity figures, not independently audited savings. The useful lesson for a smaller company is the workflow shape: find relevant knowledge, prepare a useful output, and route work to the right person.
For the proposal scenario, start with a draft-only skill. It checks whether required information is present, compares the draft with approved source material, identifies unsupported commitments, and prepares a review packet. The manager remains responsible for commercial judgment and release.
Write a skill contract before writing a longer prompt
A skill contract is Delphi's practical design recommendation, not a vendor certification. It records what the skill is for, what evidence it may use, what a good output looks like, and when it must stop. It should be short enough for the workflow owner to review without a technical interpreter.
For a proposal-review skill, define the following before connecting tools or expanding access.
- Owner and trigger: who maintains the skill, and which request starts the work
- Inputs: the proposal draft, customer requirements, and approved reference sources
- Source rules: document owners, effective dates, access restrictions, and conflict resolution
- Output: a review packet with source references, missing information, exceptions, and a draft recommendation
- Authority: prepare only; no sending, discount approval, contract acceptance, or system changes
- Stop conditions: missing current pricing, conflicting policy, restricted information, or an unsupported commitment
- Release record: skill version, source versions, test results, reviewer, and approval date
Test the awkward cases, not just the demonstration
A polished demonstration usually has complete inputs and an obvious answer. Real work has outdated attachments, contradictory instructions, missing facts, and urgent requests to make an exception. The acceptance test must cover that work too.
Build a small set of representative, appropriately protected examples with the workflow owner. Include an ordinary proposal, an expired pricing sheet, a request outside policy, a missing customer requirement, and a document containing instructions that contradict the approved workflow. Treat supplied documents as evidence, not as authority to change the agent's rules.
For each case, specify the expected behavior before running it. Sometimes the correct output is a draft. Sometimes it is an explicit refusal to complete a step until a person resolves the ambiguity. A skill that pauses for the right reason can be more useful than one that always produces a finished-looking answer.
Repeat the relevant tests when the skill, underlying model, tool permissions, or authoritative sources change. Keep the last approved version available and make it possible to disable the workflow. A shared skill is an operating asset with a maintenance obligation, not a prompt someone writes once.
The first-month operator playbook
Begin with one recurring workflow and one accountable manager. Choose work that creates enough friction to matter, but is bounded enough to run in draft-only mode. Bring the experienced employee into design as the knowledge owner, not merely as the person expected to fix every output.
Measure the complete path to an accepted result. Include source preparation, agent usage, human review, rework, and support. Faster drafting is useful only if the review queue and exception burden do not absorb the gain.
- Week 1: observe the workflow, record recurring decisions, and establish a baseline for completion and review time
- Week 2: identify authoritative sources and write the skill contract with the owner
- Week 3: test normal and exception cases without external actions; correct source and instruction gaps
- Week 4: run a limited employee pilot with human approval, version records, and a clear stop mechanism
- At review: compare accepted outputs, total handling time, rework, exceptions, and fully loaded cost with the baseline
Easy to share is not the same as ready to operate
Y Combinator's Fall 2026 request for a cloud for small software makes a related point: purpose-built tools are becoming easier to create, while deployment, permissions, and secure sharing remain hard. That is an investor's thesis, not a measured adoption result. It is still a useful warning for managers distributing AI skills across teams.
Before a local success becomes a company standard, ask who owns it, who may use it, what information it exposes, and who can stop it. Wider distribution should follow demonstrated reliability and clear support, not enthusiasm alone.
Every company needs a Company Brain, and every employee needs an AI Agent connected to the context of their role. The practical first step is not to automate everything your best employee knows. It is to turn one recurring piece of that know-how into a bounded, tested skill that helps the whole team prepare better work.
