Case studies

AI implementation evidence, not demo theatre.

Two anonymised engagements, walked through end-to-end. Five agents you can step through without us scheduling a call. Enough to size the work before you ever pick up the phone.

Case 01 · Recruitment

Reclaiming the consultant day at a Northern-England recruitment agency.

4-week build, 8 weeks of measurement. What was broken, what we built, the cadence we shipped to, and the model behind the numbers.

92%

Note completeness

11h

Median follow-up latency

60

Manager-reviewed calls / week

18%

Consultant time on admin

What was broken

Consultants on this floor were losing roughly 38% of the working day to admin: call notes typed after the fact, CRM hygiene that drifted by Wednesday, and follow-ups that quietly slipped past their SLA. The agency knew their top billers had a sharper playbook than the rest of the floor — they just couldn't see it. Placements were also leaking after start date because aftercare cadence was a personal habit, not a system.

Sector Permanent + contract recruitmentHeadcount 80 consultants · 4 desksRegion Northern England, UKStack Bullhorn · RingCentral · Outlook · LinkedIn Recruiter

Diagnosis · 2 weeks

One of our engineers spent two weeks on the floor — sitting with three desks, measuring how work moved through Bullhorn and RingCentral, and mapping the consultant day across twelve repeatable jobs. The patterns showed up fast.

70% of call notes missed the points that mattered

Managers wanted seven things captured on every call — motivation, salary, notice, location, current process, fit signals, next step. Most calls captured three or four.

Median follow-up latency on warm candidates was 4.3 days

Warm candidates routinely went cold while consultants juggled live placements. Nobody owned the nudge.

Aftercare ran on personal habit, not a system

Two desks ran day-7 and day-30 check-ins. The other two did not. Placement drop-outs were 2.4× higher on the desks without aftercare cadence.

Coaching loops were starved of evidence

Managers reviewed roughly twelve calls per week across eighty consultants. Sampling was too sparse to drive meaningful coaching.

What we built

Agents beside the live stack.

Four agents were designed to sit alongside the existing stack — no platform migration, no new logins for consultants. Each agent had a narrow remit, a defined source of truth, and an explicit human-in-the-loop checkpoint.

  • BullhornCRM of record
  • RingCentralTelephony + recording
  • OutlookEmail + calendar
  • LinkedIn RecruiterCandidate sourcing
  • SlackInternal nudges
  • Arkwright AgentsTies the agents together

Agent 01

Call Prep Agent

Triggered fifteen minutes before any calendared call. Pulls candidate or client history from Bullhorn, recent LinkedIn activity, and any open process notes. Delivers a one-page brief into Outlook with the qualification questions still outstanding.

Agent 02

Call Intelligence Agent

Listens to the RingCentral transcript, summarises the call against the seven-point checklist, drafts the CRM updates, and surfaces coaching moments to the manager queue. The consultant approves the update inside Bullhorn rather than typing it.

Agent 03

Pipeline Hygiene Agent

Watches Bullhorn for overdue actions, stale candidate status, and missing next-step dates. Sends contextual nudges into Slack — quiet by default, escalating only when a record crosses the SLA boundary.

Agent 04

Aftercare Agent

Runs day-7, day-30, and day-90 candidate and client check-ins on placed roles. Captures responses, flags risk signals to the consultant, and books a manager intervention if a placement looks fragile.

Implementation · 4 weeks of build · 8 weeks of measurement

One measurable slice at a time.

A tight build cadence kept the floor on the live system throughout. Each week we put one agent live quietly — running in the background — before it appeared in front of any consultant.

  1. Week 1

    Call Prep + Bullhorn writeback

    • Bullhorn API access + RingCentral webhook scoped
    • Call Prep Agent shipped into Outlook on six pilot consultants
    • Manager dashboard baseline established
  2. Week 2

    Call Intelligence

    • Seven-point call checklist agreed with senior consultants
    • Transcript-to-CRM drafting running quietly in the background
    • Approval UX prototyped inside Bullhorn note view
  3. Week 3

    Pipeline Hygiene + nudges

    • SLA matrix agreed with desk leads (warm 48h, hot 24h, placed 4h)
    • Slack nudge channel launched per desk, quiet hours respected
    • Manager exception report wired to weekly review
  4. Week 4

    Aftercare + review

    • Day-7, day-30, day-90 cadence deployed across two pilot desks
    • Risk-signal taxonomy reviewed with managers
    • Evidence pack assembled for the rollout decision

Results and ROI

Outcomes, with the model attached.

The tiles are the measured change. The ROI table keeps every assumption visible — pressure-test it against your own baseline, not ours.

Modelled92%Note completenessfrom 41%, across all seven checklist points
Modelled11hMedian follow-up latencyfrom 4.3 days on warm candidates
Modelled60Manager-reviewed calls / weekfrom 12, each with a scored summary
Modelled18%Consultant time on adminfrom 38% measured during diagnosis
  • Consultant time recovered

    £180k-260k / yr

    20% recovered admin time across 80 heads, valued at the loaded consultant cost band below.

  • Placements rescued via aftercare

    £90k-140k / yr

    Two additional placements retained per month, valued at the average permanent fee band below.

  • Manager coaching capacity

    Not monetised

    Coaching surface expands from 12 to 60 reviewed calls per week. Tracked as a leading indicator, not a cash line.

  • Pipeline leakage avoided

    £40k-80k / yr

    Recovered warm-candidate placements based on the latency delta and historical conversion rate from warm to placed.

Case 02 · Real estate

Tightening the viewing-to-vendor loop at a London and Home Counties estate agency.

6-week build, 10 weeks of measurement. What was broken, what we built, the cadence we shipped to, and the model behind the numbers.

94%

Feedback reaching the vendor record

7 days

Median vendor update cadence

12%

Listings missing a hero shot

+9pp

Instruction-to-offer conversion

What was broken

Negotiators ran twelve to twenty viewings a week and dictated feedback into a tool that rarely surfaced cleanly to the vendor. Vendor reports were polite, vague, and late — leaving instructions to drift toward re-marketing without a clear narrative. Listing photography varied wildly between branches because each negotiator handled their own uploads, with no shared bar for quality.

Sector Residential sales + lettingsHeadcount 180 negotiators · 22 branchesRegion London + Home Counties, UKStack Reapit · Mio · Rightmove · in-branch dictation

Diagnosis · 3 weeks

The team measured two months of viewings across six branches, listened to 140 dictated feedback files, and sat in on three vendor-update calls per branch. The same patterns came up in the calls and in the listings.

58% of dictated feedback never reached the vendor

Dictations sat in a queue waiting for a negotiator to retype them. Anything older than 48 hours was usually skipped in favour of 'I'll mention it on the call.'

Vendor calls were averaging 19 days apart

The stated cadence was weekly. The actual cadence depended on negotiator capacity, with the average drifting close to three weeks during peak periods.

34% of Rightmove listings lacked a hero shot

A hero shot here means: correctly exposed, level horizon, no parked car in frame, taken in usable light. Branch quality varied by an order of magnitude.

Re-marketing decisions had no narrative

Vendors were being asked to drop price or change agent based on 'the market' rather than concrete viewing-by-viewing feedback. Trust eroded accordingly.

What we built

Agents beside the live stack.

Three agents were scoped to sit between the negotiator and the vendor. Each one absorbed an existing manual step rather than adding a new surface to learn.

  • ReapitCRM + property record
  • MioChain tracking
  • RightmovePortal of record
  • Dictation toolViewing feedback capture
  • OutlookVendor communications
  • Arkwright AgentsTies the agents together

Agent 01

Feedback Capture Agent

Receives dictated viewing feedback from the existing tool, structures it against a five-field schema (price, condition, location, layout, intent), and posts a draft into the Reapit property record within fifteen minutes of the viewing ending.

Agent 02

Vendor Narrative Agent

Aggregates the structured feedback from the last fortnight into a vendor-ready narrative. Highlights consistent themes, flags anything that contradicts the asking price, and drafts the vendor update for negotiator approval before send.

Agent 03

Listing Quality Agent

Checks every new Rightmove listing against an agreed quality bar before it goes live. Flags missing hero shots, poor exposure, or empty descriptions to the branch lead with a recommended fix. Never auto-publishes.

Implementation · 6 weeks of build · 10 weeks of measurement

One measurable slice at a time.

A longer build window than the recruitment case reflected the breadth of data sources and the need to onboard branch leads as the human approval layer.

  1. Weeks 1-2

    Feedback capture pilot

    • Dictation tool integration scoped + five-field schema agreed
    • Two pilot branches enrolled, baseline measured on 60 viewings
    • Reapit writeback approval flow built
  2. Weeks 3-4

    Vendor narrative + cadence

    • Vendor update template reviewed with three senior negotiators
    • Cadence engine built (weekly default, escalating on stale instructions)
    • Ran quietly on 40 active instructions before negotiator approval went live
  3. Weeks 5-6

    Listing quality + rollout

    • Rightmove listing quality bar agreed with marketing lead
    • Listing Quality Agent deployed to all 22 branches in advisory mode
    • Branch-lead dashboard shipped with weekly quality scorecard
  4. Weeks 7-10

    Measurement + review

    • Vendor call cadence tracked against target across all pilot branches
    • Re-marketing decisions reviewed with management against narrative quality
    • Evidence pack assembled before the network-wide rollout decision

Results and ROI

Outcomes, with the model attached.

The tiles are the measured change. The ROI table keeps every assumption visible — pressure-test it against your own baseline, not ours.

Modelled94%Feedback reaching the vendor recordfrom 42% within 48 hours of viewing
Modelled7 daysMedian vendor update cadencefrom 19 days during peak
Modelled12%Listings missing a hero shotfrom 34% across the network
Modelled+9ppInstruction-to-offer conversiondirectional uplift on pilot branches
  • Negotiator time recovered

    £260k-360k / yr

    ~5 hours per negotiator per week recovered from feedback retyping and vendor admin, across 180 heads at the loaded cost band below.

  • Retained instructions

    £420k-720k / yr

    Tighter vendor narrative correlates with fewer instruction withdrawals; modelled at 1.5-2.5 retained instructions per branch per quarter at the average commission band.

  • Listing quality uplift

    Tracked, not monetised

    Hero-shot compliance is a leading indicator of click-through on the portal. We track it; we do not attribute revenue to it directly without a longer measurement window.

  • Manager intervention cost avoided

    £60k-110k / yr

    Branch leads previously fielded escalations from disappointed vendors. Modelled as 1-2 hours per week per branch lead, recovered at the loaded leadership cost band.

Walkthroughs

See how an agent actually behaves — without booking a demo.

Five static walkthroughs of agents we’ve already built. Pick one and step through it: what it reads, where it pauses for a person, what it decides, and what happens when it isn’t sure. You get to size the work yourself — no live demo, no compute burned, no sales pressure.

Walking through

Voice qualification

Telephony
CRM
Calendar
Email

Chapter 01

Map the workflow before anything ships.

System in scope

Telephony

Mapped during discovery, before any automation path is proposed.

System in scope

CRM

Mapped during discovery, before any automation path is proposed.

System in scope

Calendar

Mapped during discovery, before any automation path is proposed.

System in scope

Email

Mapped during discovery, before any automation path is proposed.

Discovery itemFindingWhy it matters

Primary workflow

Voice qualification

A realtime voice agent qualifies the call against an agreed scorecard and produces a CRM-ready record the consultant can accept, edit, or reject.

Systems mapped

4 connected surfaces

Telephony / CRM / Calendar / Email

Human checkpoint

Approval before commercial action

Where confidence scores, audit trail, and handoff rules sit before anything is written back live.

Want hands-on?

The live versions run for clients and scoping conversations.

Each walkthrough above is the same agent we’d build for you. When you’re ready to see one running on your data, we’ll open the live version and walk it with you.

Open a scoping call

Methodology

Composite profiles. Real implementation patterns.

We’d rather show you a faithful composite than a doctored named-client logo. Here’s what that means in practice — and the three things we hold ourselves to on every engagement.

A person stays in charge

Anything commercial is approved or reviewed by your team before it touches a live system.

Baseline before build

We measure your current workflow before we propose a single automation. No baseline, no business case.

Your stack stays put

Agents sit alongside the tools your team already uses. No migration, no new logins, no rip-and-replace.

Next step

Scope the proof point closest to your operation.

Bring one workflow, one team, and the systems already in play. We’ll work out whether there’s enough signal to build — and what a first slice should look like.