AdTech · named client · one case, one stealable principle

An ad agency became the market’s aggregator — and two weeks of planning became three hours.

Five years inside the platform rebuild that turned the UK’s largest out-of-home agency into the market’s aggregator — a DSP with its own demand and supply sides. Campaign planning fell from two weeks to three hours, and the traders stopped overriding the machine. The growth its 2017 investor called “exceptional” on exit is on the public record; the next investor named the technology as the reason it bought in.

2 wks → 3 hrscampaign planning, once traders bet on the model — measured inside the engagement
<5% → 100%of Talon’s UK bookings through Plato in one year — Talon’s CPO, on the record
6 systemsone connected platform — trading, audience data, creative management, play reporting, media-owner SaaS and white-label franchisor booking · three of them public as Plato, Ada, Atlas
2:00 · the case, narrated · rather read? — the 60-second slip is next synthetic voice · my words, my facts

Reviewing? Start here

The case in sixty seconds.

The deep read is below — every line links into it.

The outcome
Campaign planning 2 wks → 3 hrs, +£69k media value per client vs trader baselines. Public record: <5% → 100% of UK bookings through the platform in a year. the evidence →
The call
The model was right and ignored. I stopped sharpening it and changed what it handed over: a ranked list became a full plan — KPIs attached, one click to customise, every edit feeding next week’s calls. the test →
My part
Lead product designer, five years on this one platform — product definition and design across six systems, in a 50+ distributed team. the seat →
Where it breaks
Argue-back — letting the expert reshape the plan — needs an accountable expert on the other side, and can’t rescue a weak model. If either fails in your product, don’t copy this. the boundary →
The screen that turned right-and-ignored into right-and-acted-on — a plan to argue with, not a list to obey · white-labelled reconstruction, synthetic data

AdTech · Talon Outdoor · Plato, Ada & Atlas · 2019–2025

The principle this case proves

Users don’t adopt the most accurate system. They adopt the one that lets them argue back.

The house belief was the industry’s belief: make the model more accurate and adoption will come. The model got more accurate. Adoption stayed at zero. This page is about the mechanism that actually moved it.

One principle · proven at £400M scale · wrong turn included

The stakes: what believing the opposite was costing Talon Outdoor.

Week after week on the trading floor of the UK’s largest out-of-home agency, an algorithm posted better plans than the traders. Week after week, the traders played their hunches anyway.

Out of Home reaches 98% of the UK population every week, yet in 2019 buying it still worked the way it had for decades: fragmented inventory across hundreds of site owners, availability discovered by phone tag, proof-of-play arriving late and on trust. Talon — the UK’s leading independent OOH agency, handling media for Omnicom, Havas and others — decided to digitise not just its screens but, in the words of its Chief Transformation Officer Josko Grljevic, “the entire end-to-end business journey.”

That partnership was a relationship somebody had to hold, and for five years a large part of it ran through me: sitting with Talon’s commercial and product leadership, with media owners, and with the agencies buying through the platform, then coming back and deciding what got built. Not a brief handed down and drawn up — the conversation and the interface were the same job. Talon’s own account of why they chose Sahaj as the engineering partner is on the record: “We believed they had the hardcore engineering skills… that partnership has built over the last three years to such an extent where we don’t treat them as an external company, they’re an extension of our business.” I was on that team for five years — the product-definition and design side of a 50+ person distributed effort.

What “transformed” means, in the customer’s own words. The cleanest way to show the before/after is the conversations an advertiser could suddenly have:

The advertiser asksBeforeAfter
How soon can my campaign start?“At least two weeks — to search, book, and dispatch adverts.”“Upload your ads; we can start the next hour.”
Can you reach pub-goers, Fridays 9 PM–1 AM?“Impossible to cherry-pick media like that.”“The data platform recommends the sites near that audience.”
Why these sites?“Trust us — we’ve done this for years.”“Here’s the data: unique views and ranking, per site.”
Can I cancel?“A few days to roll everything back.”“Pause or cancel in one click.”
Change my ad when it rains?“Not possible.”“Yes — creative can switch on live conditions.”
Can I monitor delivery?“You’ll get a site list after the campaign.”“Views and demographics per site, as it runs.”

The ground truth: what every actor in this market actually needed.

I was the technical side of the room where the business was decided. Before anything was designed I mapped the market’s ecosystem: buyers from mom-and-pop stores to enterprises and franchisors; sellers from individual site owners to media and creative agencies, DSPs, SSPs and exchanges. Competitor teardown across the category — Google Ads, Facebook Ads, AdQuick, Hivestack, Place Exchange, Campsite (since acquired by Broadsign), among others. Then research by actor: business-model canvases, personas and journey maps with the sales team, campaign planners, media sellers, creative agencies, and the admins who keep it running. The pain-point ledgers below are drawn directly from those sessions — their words became the feature map.

Planners said
“Clients change the brief constantly; prices and availability drift under us.” · “Cherry-picking good frames in unknown areas is guesswork.” · “A 500-site campaign is unmanageable across markets.” · “Sellers don’t respond with rates in time.”
So the platform
Holds briefs, rates and bookings in one place · scores every site against audiences and recommends · runs amendment and cancellation as first-class workflows, not favours · puts sellers on notified, contract-bound response loops · generates per-site delivery reports clients can hold.
Sellers said
“Every unsold week is a permanent loss.” · “Managing 500 to 50,000 sites — prices, availability, contracts — needs a sales team we can’t afford.” · “Creatives arrive at the last moment.”
So the platform
Gave them inventory management free — heat maps of unsold media, forecasts, automated sales — and in return their inventory came onto the aggregator. That trade is the supply-side engine: the free SaaS wasn’t generosity, it was how supply showed up without a single phone call.

Why aggregation was hard enough to be worth five years. Hundreds of site owners, each with their own inventory format — some had APIs, some shared files, some just answered the phone. Sizes, renting plans and per-site rules that resist standardisation; advertising law that changes by region; availability that shifts while you’re planning against it. And the constraint that shaped my work most: adoption. Advertisers were used to choosing sites by hand — trust in automated delivery doesn’t come with the software, it has to be designed for. That last constraint is the whole next section.

The test began with me failing it: two weeks spent sharpening a model that was already sharp.

The recommendation engine outperformed the buyers, visibly. Adoption sat near zero anyway. I got the diagnosis wrong first — I spent two weeks trying to make the model smarter before admitting it was the wrong suspect. The miss stays on the page on purpose: show the miss is a house rule here, not a confession. The traders weren’t being stubborn; they were being rational. A bare recommendation asks for faith, and traders deal in collateral.

The model didn’t need to be more right — it needed to give the traders something to bet on.

The design move

So the bet became: move adoption without touching the model at all.

THE DESIGN MOVE The model didn’t change. What it handed over did. WHAT THE ENGINE PRODUCED A ranked list of billboards correct, and impossible to defend in a room no case, no cost of being wrong, nothing to adjust traders read it, then played their hunches the screen between WHAT THE TRADER RECEIVED A full campaign plan · reach, budget, formats, locations · the KPIs it will be judged on, attached · the reasoning on the card — which audiences, which environments, what trade-off Customise in one click disagreement is a control, not a dead end Every edit logged and fed into next week’s recommendations fighting the model turns into coaching it No confidence score ever reached a trader. The unit of trust was a plan they could argue with.

The design move

The model didn’t change. What it handed over did.

  1. What the engine produced

    A ranked list of billboards

    correct, and impossible to defend in a room

    no case, no cost of being wrong, nothing to adjust

    traders read it, then played their hunches

  2. What the trader received

    A full campaign plan

    reach, budget, formats, locations · the KPIs it will be judged on, attached

    the reasoning on the card — which audiences, which environments, what trade-off

the screen between

  1. Customise in one click

    disagreement is a control, not a dead end

  2. Every edit logged

    and fed into next week’s recommendations

fighting the model turns into coaching it

No confidence score ever reached a trader. The unit of trust was a plan they could argue with.

The model didn’t change — what it handed the trader did
The loop IS the principle: disagreement is fuel, not friction. A system that punishes pushback never earns a second bet.
Plate 03 · Campaign planning — reconstruction · all names and figures synthetic Redrawn from memory at production fidelity. A plan instead of a list, the KPIs on the card, and customisation the model learns from — the screen that turned right-and-ignored into right-and-acted-on.

The whole case fits in one card. Pick a planning priority and watch the plan re-shape — then open the reasoning, then customise it:

Reconstruction — anonymised, rebuilt from memory for illustration · client under NDA · figures synthetic
Recommended plan · Campaign 7

Evening CTV + roadside D6 cluster — weighted for maximum audience.

reach 1.8M est. ROI 2.1x spend 100%
Audience fit — movement patterns near sites
30-day site performance
Inventory price trend

Customised — logged · feeds next week’s recommendations

Keyboard-friendly · nothing you click here leaves the page
The strongest objection, kept in

“Letting traders reshape the machine’s plan undermines it.” That was the room’s position, and it is the reasonable one. It is also backwards: an unchallengeable system doesn’t read as authoritative — it reads as unaccountable, and experts decline to stake their name on the unaccountable. The objection dissolved the first time a buyer watched their own override sharpen the following week’s plan.

I built three things. A recommendation that arrived as a plan, not a pick. KPIs on the plan itself, so the ROI case was visible before anyone committed budget. And customisation that took one click — every adjustment logged and fed into next week’s recommendations. The first time a buyer watched their own pushback sharpen the following week’s calls, the relationship flipped from fighting the model to coaching it.

Adoption unlocked the bigger reframe: buying moments, not spots. The old brief named a place — “a billboard on Oxford Street for a month.” Once the plans were believed, the brief could name an audience and a moment — “reach party-lovers on Friday nights, 9 PM to 1 AM” — and the system chose the sites. Talon’s product strategy director described the same capability to Digital Bulletin: outcome-based measurement, omnichannel planning, behavioural audiences.

WHAT TRUST UNLOCKED The brief itself changed shape. BEFORE — THE BRIEF NAMED A PLACE “A billboard on Oxford Street for a month.” the buyer picked the sites; the system priced them once the plans were believed AFTER — THE BRIEF NAMED AN AUDIENCE AND A MOMENT “Reach party-lovers on Friday nights, 9 PM to 1 AM.” the system chose the sites; the buyer judged the case Adoption was not the end of the work — it was the thing that made a better question askable.

WHAT TRUST UNLOCKED

The brief itself changed shape.

  1. BEFORE — THE BRIEF NAMED A PLACE

    “A billboard on Oxford Street for a month.”

    the buyer picked the sites; the system priced them

  2. AFTER — THE BRIEF NAMED AN AUDIENCE AND A MOMENT

    “Reach party-lovers on Friday nights, 9 PM to 1 AM.”

    the system chose the sites; the buyer judged the case

once the plans were believed

Adoption was not the end of the work — it was the thing that made a better question askable.

What adoption unlocked — a brief that names an audience, not an address
What the trader received
Not a summary — a granular plan with every media placement in it, priced and reasoned, at the level a trader would otherwise have assembled by hand.
What they could change
Redistribute the media across the plan · remove placements they didn’t want · or rewrite the brief and run a fresh fetch. Three ways to disagree, none of them a dead end.
How the loop was judged
Not by a notification telling a trader their override had been heard — by the numbers on the next plan: impressions, reach, cover and frequency. The improvement was quantitative, which is the only kind a media buyer argues with.

The mechanism at scale

The mechanism at scale: six systems, named and public — the market, re-plumbed.

The structural play: a media agency became the market’s aggregator. The platform positioned Talon as a DSP with its own demand side — advertisers and agencies planning in one system — and its own supply side, where media owners’ inventory flowed in through free software instead of phone calls. Six systems, one connected platform — the trading floor, the audience intelligence, the creative management, the play reporting, the supply-side SaaS, and a white-label booking product franchisors brand as their own — plus PlanIt on top of the data they generated. That is what replaced the Excel sheets — not a tool bolted onto the old market, but the market, re-plumbed.

Plato — the automated trading platform: explore, plan, check availability, reserve and trade paper and digital inventory across markets, in one system. Talon’s Chief Product Officer Amy Horton, on the record: bookings through Plato went from under 5% to 100% of UK bookings in about a year, and an availability search that took three days — “48 sheets in Sheffield in June” — now takes 30 seconds. Grljevic: the MVP was built “in four months, under immense pressure.” Plato on Talon’s site ↗

Ada — the data-management platform: billions of audience data points profiling how people move, so billboards could be planned against audiences rather than addresses. McDonald’s was a launch partner; Google a regular user. Talon’s public claim for Ada-optimised campaigns: +53% consideration versus standard OOH distribution. Built, per the article, “a complete data management platform in a year, while transitioning the engineering from an incumbent.” Talon × Ada on YouTube ↗

Atlas — the programmatic buying platform, launched September 2020: audience-led, transparent, clients pay only for confirmed impressions. Grljevic’s build story is the one he showcased: existing Plato and Ada services extended into “a brand new DOOH programmatic platform in six months… we put revenue through it a month later — that is unheard of.” It has since been extended to the US market with Place Exchange. Atlas launch, ExchangeWire ↗

PlanIt — and one more product on top of the six, a client project built on the platform’s own footfall data (London Underground + National Rail): a planning tool scoring places by quietness, 2021, now deprecated but archived. Honest scale: ~1,000 monthly visitors, 250 active users. PlanIt, archived ↗

The four systems that made the claim hold — creative management: versions, formats and the rules for swapping them, so a brief that named a moment could carry creative that matched it. Play reporting: proof of play turned into reporting, which is what let the platform sell on confirmed impressions rather than booked ones. White-label booking: a franchisor brands it as its own — one central negotiation, then franchisees browse a map and book themselves, a third audience the platform had never served directly. And free inventory software for media owners: they list what they have, supply arrives without a phone call — the quiet product-led-growth mechanism under the whole aggregator.

One design system made six products one platform. I introduced a new design system across the line — a modern face, but the point was family resemblance: Plato, Ada and Atlas stopped reading as separate tools someone had bolted together and started reading as rooms of the same building. For a company selling an integrated platform, looking integrated is not cosmetics; it is the claim, made visible.

Then the design system learned to build itself. I packaged its rules as a Claude plugin — an installable artifact developers add to their own environment, not a prompt I kept re-pasting — and shipped it across the engineering org; a second plugin followed at Sahaj. Front-end work now starts on-system instead of drifting from it. Front-end work now starts on-system instead of drifting from it, and developers began prototyping with Claude themselves rather than queueing for a designer. The same argue-back principle, one level down: give the machine your rules, and it hands you back speed. Both plugins sit behind a client wall, so this one travels as a claim rather than an artifact — and without numbers, because I never measured it.

Falsifiable evidence — my numbers and theirs, kept apart on purpose.

Measured inside the engagement — mine to defend
Campaign planning 2 wks → 3 hrs once traders bet on the recommendations · £69,000 average media-value uplift per client · 45% reduction in time and effort to plan and book · 3× purchase-intent and 70% audience uplift vs traditional bookings. Not in the article; measured in the work; I’ll defend each in the room.
On the public record — Talon’s leaders, to the press
UK bookings through Plato <5% → 100% in a year · availability search 3 days → 30 seconds · Ada-optimised campaigns +53% consideration · Plato MVP in 4 months, the data platform in a year, Atlas in 6 months with revenue a month later · “the things that would normally take us 12–18 months to build were built in six months.”
The investors’ ledger — the win a standup aims for
Mayfair Equity Partners backed Talon in 2017; in July 2022 it exited fully to Equistone, calling five years of “exceptional growth”: 200+ employees across seven cities, four acquisitions, the largest global OOH planning-and-buying network across 100 markets, Apple and McDonald’s on the client list. Equistone’s stated reason for buying in: to “accelerate the investment into talent and technology” — the platforms this case describes. Deal terms undisclosed, so no multiple is claimed here; the receipts are the exit itself, and who bought, and why.
Not mine
The bidding model and the 50M-bids/hour infrastructure — engineering’s win, and they own it. The platform build speed Grljevic praises is the whole 50+ team’s, Sahaj’s engineers above all.

Where the principle breaks.

Argue-back needs an accountable expert on the other side — a trader with a number against their name, an analyst signing a memo. Casual users generate noise, not coaching; a feedback loop trained on noise degrades. And the principle cannot rescue a weak model: collateral only works when the house can pay. If either condition fails in your product, this case argues against copying it — which is exactly what a principle worth stealing should do.

The public record — watch it, read it, click it.

Start with the film: Digital Bulletin’s cover-story documentary, with Talon’s transformation chief, product director and CPO on camera about the platforms this case describes.

The film · Digital Bulletin / NODE · 2021“Talon Outdoor: An Out of Home evolution” — the client, on camera, about the transformation. Not produced by me; that’s what makes it evidence. Read the Digital Bulletin cover feature ↗
Atlas, as sold today ↗Talon’s live product page — planner, dataset mapping and studio interface shots. The current generation of the platform this case seeded. Plato, as sold today ↗The trading platform’s live page — explore, plan, reserve and trade, still the end-to-end system. The full technology suite ↗Ada, Plato and Atlas on one page, in the client’s own words.

No public walkthrough video of Plato, Ada or Atlas exists — verified again before writing this line, so none is faked here. The product-and-design evidence is above: the campaign-planner reconstruction, rebuilt from memory at production fidelity with synthetic figures, and interactive so you can work the screen a trader worked.

The rest of the record — every link verified working on 13 August 2026.

Over the past four years at Talon, Arpit has been instrumental in shaping four distinct products from the ground up. His user-focused designs are remarkably intuitive yet adept at handling complex workflows… If you need a designer who excels at combining strategic vision with practical execution, Arpit is the person to call.
Anant EastCTO at Talon Outdoor · the client, on this platform

Where the line sits

The article is Talon’s story and Sahaj’s engineering — it does not name me, and I won’t pretend otherwise. What it proves is that the platforms, the timelines and the outcomes I describe are real and were worth a magazine cover. Who held the product-definition and design pen across those five years is a question for the people who were in the room — and a named reference for this work is available on request.

Next case study → Send me the role ↗