Use casesA global consultancy

Halden Reed: every engagement started where the last one finished.

Halden Reed runs 40,000 people and thousands of engagements a year, and its agents relearned the firm's method on every one. The AI lab ran memco first, as client zero, on its own fleet of fourteen agents. Then the data engineering practice, then the first client programme. The walls held at every step.

Organisation
Global consultancy, 40,000 people in 60 countries
Client zero
The firm's AI lab: fourteen internal agents on Copilot Studio with measured baselines, eight weeks
Agents in use
Copilot Studio agents across the firm; Claude Code and Codex in engineering delivery; Claude and ChatGPT in practices
Tools connected
Microsoft 365, Confluence, Jira, GitHub
Memory networks
Firm, service line, industry, client, engagement, personal
Residency
Dedicated tenants in the EU, US and Singapore. Nothing installed at any client.
Time to first proven value
Day nine in the lab; day three on the first client programme

01Starting point

Where Halden Reed started

By 2026 every Halden Reed engagement used agents. They drafted interview guides, built models, wrote deliverables, cleaned client data and, in the engineering practice, wrote code. The firm had fourteen internal agents on its Microsoft estate with baselines for each, and a pipeline of more. Each engagement set up its own prompts, its own templates and its own rules about what the client allowed. The firm's method sat in a Confluence space few people read and no agent used.

When an engagement closed, its learning closed with it. A partner's correction to a pricing model in Madrid never reached the same practice in Chicago. The diligence report template was rebuilt on every deal because the agent did not know it existed. The head of the AI lab counted the first fortnight of a new engagement: a week of a junior consultant teaching the agents the house style, and a partner reviewing first drafts that repeated mistakes corrected the month before.

The obvious fix was the dangerous one. One firm-wide memory puts a client's figures within reach of a team serving its competitor. The firm's requirements were written down before any vendor was spoken to: strict client separation, a controlled way to generalise a lesson across client and industry lines, a point-in-time audit for regulated work, a concrete path from a human correction to a reusable lesson, integration beneath the tools the firm already used, scopes by firm, service, industry, client and engagement, and nothing installed at the client. The lab would prove it internally before any client heard the word.

02Day one

Day one

Client zero was the firm itself. The lab connected its fourteen agents in a week against baselines it already had, and spent the second week with risk on the one rule that mattered: how a lesson moves up from an engagement without carrying the client with it.

  • Copilot Studio agentsMCP connection for the lab's fourteen internal agents, knowledge-work domain, under the firm's identity. Each agent's existing performance baseline became the pathfinder's comparison.
  • Claude Code and Codexmemco plugin for the engineering delivery practice, added in the practitioner validation stage.
  • Confluence and JiraThe method library and engagement tickets link to the insights that cite them, with page version and date.
  • PractitionersWhen a partner or manager corrects a draft, the agent proposes the lesson with the client removed, and the reviewer saves it to the engagement, offers it upward, or skips it.
  • The abstraction gateA lesson offered above an engagement is rewritten as a general rule, reviewed by the service line's knowledge lead, and promoted carrying no client content, figure or permission.
  • TenantsThree dedicated tenants, in the EU, US and Singapore, administered by the firm's knowledge function. No software at any client.
  1. FirmThe method: deliverable standards, house style, what never enters memory, regulated-work conduct.Everyone, every agent.
  2. Service lineA practice's accepted patterns, templates, benchmarks and standing corrections.Practitioners in the service line.
  3. IndustrySector rules that hold across clients: regulatory calendars, benchmark sets, what a sector's boards expect.Practitioners tagged to the industry.
  4. ClientA client's standing facts: fiscal year, approval chain, what it will and will not accept.People on that client's engagements, as the firm's conflicts system allows.
  5. EngagementThe work: findings, figures, decisions, what the client said. Sealed.The engagement team only. Never read sideways, promoted only through the gate.

fig. 01Scope map

Halden Reed memory networksNested scopes: the firm network holding the method, service line networks inside it, industry networks inside that, client networks, and sealed engagement networks at the centre; the only route up is the abstraction gate, a rewritten general rule approved by a service line's knowledge lead.FIRMthe method; three tenantsSERVICE LINEdata engineering, focalINDUSTRYretail, banking, healthCLIENT NETWORKSas conflicts allowENGAGEMENT NETWORKS · SEALEDEngagement 2041Engagement 2058+ one per engagementABSTRACTION GATEKNOWLEDGE LEADReads inherit downward · Writes stay local · Engagements never read sideways
Halden Reed memory networksNested scopes: the firm network holding the method, service line networks inside it, industry networks inside that, client networks, and sealed engagement networks at the centre; the only route up is the abstraction gate, a rewritten general rule approved by a service line's knowledge lead.FIRMthe method; three tenantsSERVICE LINEdata engineering, focalINDUSTRYretail, banking, healthCLIENT NETWORKSENGAGEMENT NETWORKS · SEALEDEngagement 2041Engagement 2058+ one per engagementABSTRACTION GATEKNOWLEDGE LEADReads inherit downward · Writes stay local · Engagements neverread sideways
Five levels in three tenants. Engagement networks are sealed. The dashed route is the abstraction gate: a lesson leaves an engagement only as a general rule with no client content, figure or permission.

03The first weeks

The first weeks

The lab's first weeks were a measured experiment: fourteen agents, their baselines, and a readout in week eight. The lessons below are the ones that shaped the firm's rules and the ones that proved the walls.

fig. 02Deployment timeline

Halden Reed: client zero to the first client programmeTimeline across two segments: eight weeks of client zero in the AI lab, from fourteen agents connected against their baselines through the first reuse on day nine, the gate holding a client figure on day eleven and the readout in week eight; then months four and five, the practice readout and the first client programme with its first reuse on day three.BASELINES IN HANDCLIENT ZERO, AI LABPRACTICE, THEN FIRST CLIENTDAY 156DAY 9First reuse, anotherpracticeDAY 153First reuse, day threeDAY -7Fourteen lab agentsconnectedDAY 0Client zero startsDAY 4Slide-title correctionDAY 11Gate holds a clientfigureDAY 56Readout; risk signsoffDAY 120Practice validationreadoutDAY 150First client programmeconnects
Halden Reed: client zero to the first client programmeTimeline across two segments: eight weeks of client zero in the AI lab, from fourteen agents connected against their baselines through the first reuse on day nine, the gate holding a client figure on day eleven and the readout in week eight; then months four and five, the practice readout and the first client programme with its first reuse on day three.Baselines in handClient zero, AI labPractice, then first clientDAY -7Fourteen lab agents connectedDAY 0Client zero startsDAY 4Slide-title correctionDAY 9First reuse, another practiceDAY 11Gate holds a client figureDAY 56Readout; risk signs offDAY 120Practice validation readoutDAY 150First client programme connectsDAY 153First reuse, day three
Two segments with a break between them. The filled marks are first reuse in the lab, day nine, and on the first client programme, day three.
  • Day 4. A manager corrects a deck for the third time that week: conclusions buried in chart titles.Lead each slide with the conclusion in the title. Charts support the title; they do not replace it.Offered to the firm network and approved by the knowledge lead. Day nine: an agent in a different practice drafts a deck that way unprompted. First reuse. The correction stops recurring across the lab's fourteen agents within a fortnight.
  • Day 7. A lab agent cites a benchmark set three years old in a retail model.Use the retail practice's benchmark set dated this quarter. Flag any range older than four quarters.Saved in the retail service line network with the current set linked. Stale benchmarks stop appearing in the lab's drafts and, later, in the practice's.
  • Day 11. A lesson from a client engagement is offered to the industry network with the client's margin figure in it.Capital adequacy models use the revised rules in force from 1 October. Models with the old treatment are flagged for rework.The gate holds it. The knowledge lead rewrites it as the general rule above, with no client, figure or model name, and promotes that. The engagement keeps its own version. The firm's rule becomes explicit: the lesson promotes, the evidence does not.
  • Day 16. A service-line insight about executive summaries is applied to a regulated client whose board requires a different format, and a partner sends it back twice.Executive summaries are one page: the answer, the three reasons and the ask. Regulated clients' board formats are client facts and live in the client network.Scoped, with the client's format saved where it belongs. The partner approves the next ten service-line insights himself before he trusts the agent's drafts again.
  • Week 8. The readout. Fourteen agents against their own baselines.The lab's agents reused 61 lessons across 9 of the 14 agents; first-draft acceptance by reviewers rose from 47 to 69 percent; a regulated-work audit reconstructed what an agent knew on a given day in under an hour.Risk signs off practitioner validation in the data engineering practice. The lab's readout, not a deck, is what the practice lead reads.
  • Month 5. The first client programme, in the data engineering practice, connects.The service line's 140 approved lessons, read from the first session, with the client's own engagement network sealed from day one.First reuse on day three instead of day nine. The engagement team spends its first week on the client's problem, not on teaching agents the house style. Nothing from any other client is readable, and the client's security team is shown the log that proves it.

04Governance

Governance

  • Engagements are sealedAn engagement network is created with the engagement, readable by its team as the conflicts system allows, and archived at close. Nothing is read across engagements, ever.
  • The gate is the only way upA lesson leaves an engagement only as a rewritten general rule, approved by a service-line knowledge lead, carrying no client content, figure, document or permission. The record keeps who offered it, who approved it and when, without exposing the source.
  • Point-in-time audit for regulated workEvery insight is versioned with when it served and when it changed. For regulated engagements, risk can reconstruct what an agent knew on a given day, which the lab proved in under an hour.
  • Internal first, client secondThe firm validated the loop on its own agents, then on one practice, before any client programme. Client programmes run in the firm's tenants with nothing installed at the client.
  • Residency, audit and retentionThree tenants in the EU, US and Singapore, each under the firm's identity. Every read, write, promotion and archive is exportable. Engagement memory is retained to the firm's records schedule and exportable to the client on request.

fig. 03The write path

From a manager's correction to every practice, through the gateSwimlane showing a manager correcting a deck for burying its conclusions, the lab agent proposing the rule, the manager offering it upward, a knowledge lead approving it for the firm network, memco serving it to every agent across three tenants, and an agent in another practice drafting that way on day nine.MANAGERLAB AGENTKNOWLEDGE LEADmemcoAGENT, ANOTHERPRACTICECorrects a deck:conclusionsburiedday 4, third timeProposes: leadeach slide withthe conclusionOffers it upwardABSTRACTION GATEApproves for thefirm networkServed to everyagent, threetenantsDrafts the deckthat wayunpromptedday 9
From a manager's correction to every practice, through the gateSwimlane showing a manager correcting a deck for burying its conclusions, the lab agent proposing the rule, the manager offering it upward, a knowledge lead approving it for the firm network, memco serving it to every agent across three tenants, and an agent in another practice drafting that way on day nine.MANAGERCorrects a deck: conclusions buriedday 4, third timeLAB AGENTProposes: lead each slide with theconclusionMANAGEROffers it upwardKNOWLEDGE LEAD · ABSTRACTION GATEApproves for the firm networkmemcoServed to every agent, three tenantsAGENT, ANOTHER PRACTICEDrafts the deck that way unpromptedday 9
The dashed node is the gate: a knowledge lead approves a rewritten general rule. The engagement keeps its own version; the firm gets the rule.

05What compounded

What compounded

The lab's readout at week eight, the practice's at month four and the first client programme's at month six, each against its own baseline.

MetricMovementWhy
Cross-person or agent learning gain61 lessons, 9 of 14 agentsBy week eight of client zero. By month six, a lesson learned in one country was being applied in another within a week, on average.
First-pass acceptance47% to 69%Reviewer acceptance of first drafts in the lab. The data engineering practice moved from 51 to 72 percent in its validation stage.
Time to first proven valueDay nine, then day threeDay nine in the lab, day three on the first client programme, which started from 140 approved service-line lessons instead of a method library nobody read.
Repeat-error rateA week to an afternoonThe first week of every engagement spent teaching agents the house style became an afternoon, measured on the first client programme as the time until a partner accepted drafts without a standing correction list.
Adaptation speedOne dayThe capital adequacy rule, promoted through the gate, was read by agents on four engagements the day it was approved.
Improvement with experienceStill climbing, three tenantsThe firm's method is now something its agents read, and it is the firm's, not any engagement's or any model's.

fig. 04Lab and practice

Reviewer acceptance of first drafts, in the lab and in the practiceSlopegraph: reviewer acceptance of first drafts rises from 47 to 69 percent across the lab's eight weeks, and from 51 to 72 in the data engineering practice's validation stage.STARTREADOUT51%47%72%69%AI lab, fourteen agentsData engineering practice
First drafts accepted by reviewers without a correction, in the AI lab over eight weeks and in the data engineering practice over its validation stage.

fig. 05Costs before it pays

Reviewer edits per hundred first drafts in the labLine chart: reviewer edits in the lab fall from 53 per hundred first drafts to 31 by week eight, with the first fortnight almost unchanged.30406053BASELINE5031WK 2WK 4WK 6WK 8Reviewer edits per 100 first drafts, AI lab
Reviewer edits per hundred first drafts across the lab's fourteen agents, against their existing baselines. Reads added six percent to tokens in the first fortnight.

06What it cost

What it cost

  • Setup timeA week to connect the lab's fourteen agents and a week with risk on the gate. Eight weeks of client zero against existing baselines. Practitioner validation in one practice over the following quarter. Six months from first connection to the first client programme.
  • Who ran itTwo people in the AI lab, two in the knowledge function and one in risk. Knowledge leads spend about two hours a week on their gates. No engagement team changed its workflow.
  • Licence and tokensA paid pathfinder for the lab, then an enterprise agreement covering the three tenants and named service lines, with client programmes added as named boundaries. Reads added six percent to the lab agents' token use; by week eight tokens per accepted draft were down a third and partner review time was down by more.
  • What it replacedThe method library in Confluence, the per-engagement prompt packs, and the first week of every engagement spent teaching the agents the house style.

Prove it on your own agents first.

Join the waitlist

More deployments.

All use cases