graymatter.
Menu

Representative projects

The patterns behind useful agents.

These are composite scenarios grounded in systems and methods Phil has built and used. They show the kind of problem, working system, control and client decision Graymatter can deliver.

They are not claims about a named client and contain no invented result, testimonial or performance number. The setting has been generalised so confidential work stays confidential.

01From market signal to accountable next actionGROWTH + DECISION SUPPORT02A review queue that turns corrections into operating evidenceQUALITY + HUMAN FEEDBACK03Executive briefs with receipts, not confident summariesKNOWLEDGE + EXECUTIVE BRIEFING04A proposal agent that stops before the promisePROPOSALS + HUMAN COMMITMENT05A service guardian that tells the truth about what is liveOPERATIONS + OBSERVABILITY06A control room that separates activity from progressPORTFOLIO + GOVERNANCE07A voice front desk that knows when to hand the call backVOICE + CUSTOMER INTAKE08An employee service agent that shows the policy behind the answerEMPLOYEE EXPERIENCE + APPROVED KNOWLEDGE
01

GROWTH + DECISION SUPPORT

From market signal to accountable next action

Representative composite · Mid-sized B2B services · Working Prototype → Pilot to Practice

Representative situation

A growth team sees acquisitions, leadership moves, funding, tenders and competitor activity every day. The signal looks useful, but public claims, the current account record and relationship history do not always agree. Acting on the wrong version can damage a real relationship.

CONTROL POINT

Research and assembly can be automated. Relationship truth and outreach stay with the account owner.

What the working system does

  1. 01

    Gather allowed public signals and retain the source

  2. 02

    Verify the organisation, event and date before promotion

  3. 03

    Reconcile the signal with the current account record, owner and activity

  4. 04

    Rank the opportunity and prepare a short brief with uncertainty visible

  5. 05

    Offer next-step options without sending or changing the relationship record

WHAT EXISTS AT THE END

A working triage prototype, source trace, review boundary and the evidence needed to decide whether a controlled pilot is worthwhile.

02

QUALITY + HUMAN FEEDBACK

A review queue that turns corrections into operating evidence

Representative composite · Shared service or operations team · Working Prototype → Ongoing AI Partner

Representative situation

An AI system produces useful recommendations, but routine and uncertain cases arrive in the same stream. Reviewers correct mistakes in chat or email, so the organisation cannot see where the system is weak or whether it is getting safer over time.

CONTROL POINT

Nothing is silently promoted because the model sounds confident.

What the working system does

  1. 01

    Route ambiguous and higher-impact outputs into a narrow review queue

  2. 02

    Show the recommendation beside the evidence used to reach it

  3. 03

    Capture the reviewer’s decision and reason in a structured record

  4. 04

    Use repeated decisions to calibrate rules, tests and thresholds

  5. 05

    Keep promoted outputs traceable to their source and approval

WHAT EXISTS AT THE END

A bounded queue, reviewer interface, durable decision record and a calibration backlog grounded in actual corrections.

03

KNOWLEDGE + EXECUTIVE BRIEFING

Executive briefs with receipts, not confident summaries

Representative composite · Executive and account leadership · Working Prototype

Representative situation

Leaders need a compact view assembled from approved websites, news, internal records and past activity. Manual preparation is slow. Generic AI summaries flatten conflicting facts, age quickly and hide what has not been checked.

CONTROL POINT

Missing evidence remains visible. A polished sentence is not treated as proof.

What the working system does

  1. 01

    Search approved internal and public sources

  2. 02

    De-duplicate repeated claims and preserve source lineage

  3. 03

    Cross-check time-sensitive and material facts

  4. 04

    Separate sourced fact, inference, recommendation and open question

  5. 05

    Produce a short brief with the evidence close to each claim

WHAT EXISTS AT THE END

A briefing workflow, approved-source map, evidence standard, test set and a clear decision about where expert review remains essential.

04

PROPOSALS + HUMAN COMMITMENT

A proposal agent that stops before the promise

Representative composite · Mid-sized professional services · Working Prototype → Pilot to Practice

Representative situation

A team repeatedly assembles first drafts from tender questions, service material, account notes and earlier proposals. The expensive work is not typing. It is finding the current approved evidence, spotting what is missing and keeping commercial commitments with the responsible person.

CONTROL POINT

The agent prepares. Named people approve every external claim, price and commitment before anything leaves the organisation.

What the working system does

  1. 01

    Read the request and create a requirements and evidence checklist

  2. 02

    Retrieve only from the approved content and current account sources

  3. 03

    Draft each response with its source and unresolved gaps visible

  4. 04

    Route price, capability, legal and delivery commitments to named owners

  5. 05

    Retain the draft, source set, reviewer changes and approval record

WHAT EXISTS AT THE END

A source-grounded first-draft prototype, review gates, acceptance tests and an honest view of the preparation work that can be removed safely.

05

OPERATIONS + OBSERVABILITY

A service guardian that tells the truth about what is live

Representative composite · Small portfolio of live AI services · Ongoing AI Partner

Representative situation

A remote AI service can appear healthy while a schedule, dependency, information path or configuration has quietly failed. A non-technical owner needs an honest status without reading infrastructure logs or mistaking a local scaffold for a live result.

CONTROL POINT

The guardian reports and recommends. It does not make destructive changes on its own.

What the working system does

  1. 01

    Run bounded, read-only health and schedule checks

  2. 02

    Separate live success from a local scaffold or stale status

  3. 03

    Classify findings as green, amber or red with supporting evidence

  4. 04

    Redact credentials and expose only the fields needed for diagnosis

  5. 05

    Escalate a specific action when human permission is required

WHAT EXISTS AT THE END

A read-only check set, clear status contract, evidence-backed escalation and a recovery path the service owner can understand.

06

PORTFOLIO + GOVERNANCE

A control room that separates activity from progress

Representative composite · Leadership team with several AI initiatives · Opportunity Map → Ongoing AI Partner

Representative situation

Leaders need to distinguish what is live, what is running in shadow mode, what has paused, what failed a quality check and what still requires approval. A single ‘active’ label makes experiments look like outcomes and recommendations look like executed actions.

CONTROL POINT

The interface does not make a recommendation look like an executed action.

What the working system does

  1. 01

    Represent production, shadow, paused and failed-quality work separately

  2. 02

    Show the owner, next decision and latest evidence for each item

  3. 03

    Keep approval-gated actions visibly blocked until approved

  4. 04

    Connect operating status to a plain-English management brief

  5. 05

    Preserve the underlying record so the summary can be challenged

WHAT EXISTS AT THE END

A readable portfolio view, explicit state model, owner and decision record, plus a backlog leadership can prioritise without AI theatre.

07

VOICE + CUSTOMER INTAKE

A voice front desk that knows when to hand the call back

Representative composite · Mid-sized service organisation · Working Prototype → Pilot to Practice

Representative situation

Routine calls arrive after hours, during peaks or with too little information for the next person to help. A voice agent can reduce waiting, but a fluent conversation becomes dangerous if it hides uncertainty, loses the caller or makes a commitment it does not own.

CONTROL POINT

The voice agent has a human exit. It does not take payment, change an account, decide eligibility or make a contractual promise.

What the working system does

  1. 01

    Identify the service and explain the automated role where required

  2. 02

    Capture the reason for the call through a short approved conversation

  3. 03

    Answer only from the bounded service and policy knowledge set

  4. 04

    Book, route or prepare the next step when the caller confirms the details

  5. 05

    Hand back to a person on uncertainty, sensitivity, distress or request

  6. 06

    Record the transcript, structured summary, outcome and hand-off state

WHAT EXISTS AT THE END

A working call flow, realistic conversation test set, latency and hand-off evidence, disclosure and retention decisions, and a clear view of which calls should never be automated.

08

EMPLOYEE EXPERIENCE + APPROVED KNOWLEDGE

An employee service agent that shows the policy behind the answer

Representative composite · Distributed mid-sized workforce · Working Prototype → Ongoing AI Partner

Representative situation

Employees ask the same practical questions across email, chat and intranet pages, while HR spends time locating the current policy and explaining who owns an exception. A generic assistant may answer quickly but cannot be trusted if it blends old and current material.

CONTROL POINT

The agent explains approved information. It does not decide pay, performance, leave exceptions, discipline or any other employment outcome.

What the working system does

  1. 01

    Search only the approved policy and employee-service sources

  2. 02

    Answer the practical question and show the relevant source passage

  3. 03

    Check the policy version, effective date and employee context allowed for the task

  4. 04

    Route exceptions and personal matters to the correct human owner

  5. 05

    Capture unanswered questions as a knowledge and service backlog

WHAT EXISTS AT THE END

A controlled knowledge service, source and version path, question test set, escalation design and evidence about which employee requests are safe to answer directly.

Why the detail is generalised

Confidential work should stay confidential.

Graymatter will not publish a client logo, system detail, internal name or performance number merely to make a portfolio look busy. Identifying material is shared only with permission.

When an approved case note is available, it will state the baseline, intervention, observed result, limitations and measurement period. Until then, these representative project patterns are the honest proof available.

A practical first conversation

Which pattern looks like your work?

Bring the workflow, the systems around it and the decision a person still needs to own. We can tell you whether one of these patterns fits or whether a simpler approach is better.

Start with the workflow