Skip to content

The workflow coach and the AI use check ​

The workflow coach is a proposed AI assistant in the employee's app. It works from the person's role and their own Task Cards. It has one goal for every Task Card: help the person do the work faster, do more of it, or do it in a way that is cheaper and safer.

WARNING

Planned, with a first part built. The client app's DutyGraph Agent (November 14, 2026) is a first, smaller version of the coach: on one of the person's own Task Cards it offers ideas to make it faster, asks for one missing part at a time, and times one run on the device. The full coach and the AI use check are not built. The role persona that the coach would use is built; see The role AI persona. The Status box at the end says which part is which. All prices and numbers on this page are examples.

The four aims ​

AimThe coach's questionExample of a proposal
FasterWhich step takes the most time?"Step 3 is a copy from one screen to another. A bot could do it."
MoreWhat stops you from doing this more often?"You wait 2 days for an approval. Ask for a standing limit."
CheaperDoes this step need the most expensive AI model?"A smaller model gives the same result on this step."
SaferDoes company data go somewhere it must not go?"This step puts customer data into a personal AI account. Use the company account."

The coach proposes one change at a time. The person decides. A changed step on a Task Card needs the same acceptance as any other change to the card.

The AI use check ​

Many employees already use AI in their work, and most companies do not know where. When a person describes how they do a task, the coach asks one more question for each step that uses AI.

What the coach asks ​

  1. Which AI tool? For example Claude, ChatGPT, Copilot or Gemini.
  2. Which model and setting? For example the largest model with long reasoning, or a small fast model.
  3. Whose account? A company account or a personal account.
  4. What goes in? For example public text, internal numbers, or customer data.

The answers are saved on the step of the Task Card as AI used.

How the answers are used ​

mermaid
flowchart TD
    A["The person describes a step<br/>and says which AI they use"] --> F["AI used, on the Task Card step<br/>tool, model, account, data that goes in"]
    F --> C{"The coach checks<br/>3 things"}
    C --> COST["Cost<br/>current model prices<br/>from a live price list"]
    C --> FIT["Fit<br/>does the step need this model<br/>and this reasoning level?"]
    C --> SAFE["Data rules<br/>the company's own AI policy<br/>and the labels on the card"]
    COST --> T["A test on the person's real task<br/>before any recommendation"]
    FIT --> T
    T --> R["Proposal to the person<br/>'Same result, lower cost'"]
    SAFE --> W["Note to the person first<br/>'Use the company account for this step'"]
    R --> D["The person decides"]
    W --> D
    classDef planned fill:#eeeeee,stroke:#888,color:#333,stroke-dasharray: 4 3;
    class A,F,C,COST,FIT,SAFE,T,R,W,D planned;

Cost: the right model for the step ​

A large model with long reasoning costs many times more than a small model. A person who uses the largest model for every step pays for capacity that most steps do not need.

Kind of stepWhat it usually needs
Sort, tag, extract a field, reformatA small, fast model with no extended reasoning
Summarize, draft a routine messageA middle model
Plan, analyze a hard case, write for a customer or a regulatorA large model, with reasoning turned up

This table is a starting point, not a rule. Three checks apply before the coach recommends a change:

  • Prices come from a live price list. The coach reads current model prices from a pricing source, for example the OpenRouter model list. It never states a price from its own memory, because prices change.
  • A test comes before a recommendation. The coach runs the person's real step on the cheaper setting and shows both results. It recommends the change only when the person agrees that the result is as good.
  • Quality is the person's call. The person can say "keep the larger model" with no reason.

Example (fictional numbers) ​

ItemNowProposed
StepSort 300 support emails a week into 6 groupsThe same
ModelLargest model, long reasoningSmall model, no extended reasoning
Cost per week$18.00$0.90
Result on a test of 50 real emails49 correct48 correct
The person's decisionAccept. The one different email goes to a person in both cases.

Safety: data and the company's rules ​

The coach compares each AI used entry with the company's own AI policy.

What the coach seesWhat it tells the person
Customer or patient data goes into a personal AI account"Use the company account for this step. A personal account is outside the company's agreement."
A password, key or account number is in a prompt"Remove it. Use the company password vault."
An AI tool that is not on the company's approved list"This tool is not approved yet. You can ask for it, or use the approved one."
A step with a decision that must be recorded"Keep the AI output with the record, so the decision can be audited."

The same four labels as in Employee Knowledge Compensation apply to what goes into an AI.

The tone: nobody is in trouble ​

People hide their AI use when they think it will be punished. A hidden use cannot be made safe. The check is designed around this:

  • The person sees it first. A data note goes to the person, with the fix. What the company sees after that follows the company's own policy, which the person can read before they answer.
  • No blame for the past. The question is "how do you do this today?" The purpose is a safe way to keep doing it.
  • Every "no" comes with a way forward. A flagged tool comes with the approved one, or a way to request it.
  • No ranking. DutyGraph does not compare people by their AI use or their AI spend.

What the company gets ​

  • A list of where AI is used, by workflow and step, from what people said. It is not from monitoring.
  • AI spend by workflow, where prices are known, and the saving from accepted changes.
  • Policy gaps: tools that people need and the company has not approved.
  • Evidence for an audit: which steps use AI, on which account, with which kind of data.

Requests from your advisor, and wait time ​

The advisor or forward-deployed engineer who runs the engagement can send a person questions and requests in the app. DutyGraph records the time from when a request is sent to when it is answered.

The coach would use this in two ways:

  • For the person: a quiet list of open requests, oldest first, with "waiting since Tuesday". No countdown and no red timer.
  • For the chain: the wait on a request is the same kind of fact as the wait between two Task Cards in Workflow Gainsharing. It shows where work stops.

Response time is a fact about a request. It is shown to the advisor who sent it. DutyGraph does not show it on the team board and does not rank people by it.

Notifications a person cannot miss ​

Two events need the person's attention: a request from the advisor, and a Task Card that another person finished and that the person's own work now depends on. One channel is not enough, because a person can miss any single channel.

At the start of an engagement, each person chooses and tests their channels. The setup step sends a test message to each one, and the person confirms that it arrived.

ChannelWhat it can doLimit
Phone push from the web appA notification on the lock screen. An Android phone can vibrateOn an iPhone, the person must first add DutyGraph to the Home Screen
Phone push from the phone appA notification with sound and vibrationThe person must install the app
EmailReaches everyone. A "priority" subject line and sender nameEasy to miss in a full inbox
Text message or WhatsAppRead quickly by most peopleThe person must agree to receive it. The company pays for each message. WhatsApp needs an approved business sender
mermaid
flowchart LR
    E["Event<br/>a request for you, or<br/>a finished card you depend on"] --> P["Push<br/>now"]
    P -- "not opened<br/>after 1 hour" --> M["Email"]
    M -- "not opened by<br/>the next work day" --> T["Text or WhatsApp<br/>if the person agreed"]
    T -- "still open" --> A["The advisor sees<br/>'not reached' and<br/>can call the person"]
    classDef planned fill:#eeeeee,stroke:#888,color:#333,stroke-dasharray: 4 3;
    class E,P,M,T,A planned;

The times in the diagram are examples. The company sets them, with quiet hours. A message never contains the answer or the content of a card. It says what is waiting and gives one link.

How the coach uses the role persona ​

The coach starts from the role AI persona: what the role watches, what it decides alone and what it sends to another role. It then reads the person's confirmed Task Cards. It can therefore say "you decide this alone" or "the Finance Manager decides this", and it does not propose a change that is outside the person's authority.

Rollout order ​

  1. The question. Add AI used to the Task Card step. The company gets the list of where AI is used.
  2. Data notes. Compare each entry with the company's AI policy and show the note to the person.
  3. Cost. Connect a live price list and show the cost of each AI step.
  4. Test and propose. Run the person's real step on another setting and show both results.
  5. The coach conversation. "Help me make this faster", with the four aims, and with Workflow Gainsharing when the company turns it on.

INFO

Status. Built: the role AI persona; requests from the advisor to a person, with response times that the advisor can list; governance checks on agent requests; the removal of secrets from Task Cards. Planned: the coach in the employee's app, the AI used field on a Task Card step, the comparison with a company AI policy, the live price list, the test on a cheaper setting, the company views of AI use and spend, the list of open requests with wait time in the employee's app, the notification when a finished Task Card makes work ready for the next person, text message and WhatsApp channels, the step from one channel to the next, and the channel test at setup. Built (Partial): email notices and web push for advisors; push to team members' phones is off unless the server turns it on. The client app's DutyGraph Agent (make a card faster, finish a card, time a run on the device, the advisor's questions) is built and connected to the company's AI connection; times are not sent to the server.

Source: gitbook/product/roadmap/workflow-coach.md in the DutyGraph repository.

DutyGraph. Pages label each feature Built, In review or Planned.