The SaaS Distribution Course #17: How Bespoke Turned Human Replies Into 150+ Public Deployments

TL;DR

Bespoke reports more than 150 deployments of its public-facing AI across airports, local governments, and national infrastructure. The system started with the opposite of scale: founder Akemi Tsunagawa stopped travellers in stations, recruited interviewees through a dating app, and sometimes had humans answer behind an early chatbot.

That manual layer produced question patterns and credible usage. A hotel use case led Narita Airport to approach the company in 2017. Narita put Bebot on its Wi-Fi landing page, removing the download problem, while JR East helped with customer acquisition and Narita supplied passenger insight. Years later, the same habit of observing customer work moved Bespoke from chat into multilingual training and then into early shipyard robotics.

The reusable mechanism is field observation → bounded manual fulfilment → institutional anchor → embedded access → partner trust transfer → adjacent-work discovery. The warning: the 150+ count is company-reported, the funnel economics are private, and the robotics stage is still development rather than a proven scaled channel.

What you will build

You will build an evidence-to-institution distribution system. It is for a product whose buyer cannot trust a landing-page promise yet. The outputs are a field-work ledger, a manual fulfilment boundary, an anchor-account proof packet, an embedded-access plan, a partner-transfer map, and an adjacency queue. Each output has a pass condition and a stop condition.

Use this system if

• the workflow is new enough that buyers cannot evaluate it from category language;
• you can deliver one narrow outcome manually without hiding material limitations;
• one credible institution can expose the product to many real users;
• usage produces structured questions, exceptions, or training data;
• existing customers repeatedly ask for the next adjacent job.

Do not use it if

• manual fulfilment would create safety, privacy, or compliance risk;
• the “anchor logo” has no distribution surface and cannot supply usable evidence;
• every customer requests a different problem with no shared data or workflow;
• you need to mislead users about who or what performs the work;
• a healthy self-serve loop already produces faster, cheaper learning.

Verified case snapshot

Stage	Evidence	Limit
2015 discovery	Traveller interviews moved the idea from a local guide towards concierge help	Interview volume and conversion are undisclosed
Manual proof	Humans answered during early chatbot tests to learn demand and dialogue	Frequency and duration are founder-reported
2017 anchor	Narita Airport launched Bebot and exposed it through airport access points	Contract value and sales-cycle length are private
Partner expansion	JR East helped customer acquisition; Narita shared passenger insight	Attribution split and partner terms are undisclosed
Current footprint	Bespoke reports 150+ deployments across public and infrastructure settings	Company-reported, not independently audited
2024 adjacency	BeTrained launched; an Ehime programme implemented it at two companies	Verified implementations, not a broad revenue claim
2026 adjacency	Robotics development began after industrial customers requested more automation	No scaled deployments or revenue disclosed

The endpoints have unusually good public support. Narita's own automation page lists BEBOT. A 2017 launch release gives a 14 November service date. The current Bespoke company page reports 150+ deployments, while a Japan Times feature independently confirms millions of users, 11 languages, and the Narita and Tokyo Station footprint. None of those sources reveals CAC, payback, margin, retention, or channel-sourced revenue.

The model: turn service evidence into a trust-transfer chain

Manual work is not the distribution system. It is the first evidence generator. The system forms only when each stage makes the next one cheaper or more credible. Direct interviews reveal the job. Manual replies reveal the vocabulary and exceptions. An anchor institution proves the workflow under real constraints. Embedded access puts the product in front of users without asking them to discover or install it. Partners transfer trust to the next account. Repeated adjacent requests determine where the product should move.

1. Observe: capture the language and existing workaround.
2. Fulfil: perform one bounded outcome and log every exception.
3. Anchor: choose a buyer whose environment creates credible proof.
4. Embed: place the workflow inside an existing user path.
5. Transfer: let partners and proof shorten the next trust decision.
6. Expand: follow repeated adjacent work, not random feature requests.

Bespoke's path is useful because the stages remain visible. The company did not begin with a chatbot thesis. It began with traveller problems. It did not discover industrial training from a market-size slide. Customers exposed it. Robotics did not arrive because “physical AI” became fashionable; shipyard and construction teams asked what else could be automated after training.

Step 1: recruit the work, not a persona

Tsunagawa went where the work was visible: Shibuya Crossing, Roppongi Station, and later a dating app that could surface non-Japanese travellers. An earlier founder interview says the first six months focused on a consumer service. User requests for reservations and basic help exposed the real job, while hotels and stations described the same language barrier from the other side.

Do this: recruit 12 people while they are close to the painful workflow. Record the last action, the workaround, the delay, the consequence, and the words used. Do not ask whether they “like” your idea.

Output: a job ledger with at least eight recent examples and three recurring exception types. Pass condition: five people independently describe the same blocked outcome. Stop condition: the pain exists only in hypothetical answers or each person needs a different product.

Step 2: operate the smallest truthful service

In 2015, investors doubted both the technology and the habit of talking to AI. Bespoke sometimes put humans behind the interface. Years later, a Christmas software failure forced Tsunagawa to answer for a week. That outage taught the team that human-like conversation improved engagement; it then copied the useful interaction patterns back into the bot.

The lesson is not “fake your product.” Define a manual boundary the buyer can accept: the exact outcome, operating hours, escalation path, data access, and evidence you will capture. If a human must read sensitive messages, say so. If the workflow concerns health, finance, safety, or legal decisions, require the relevant controls before running it.

Metric: track successful outcomes, unanswered intents, response time, repeat usage, and corrections per 100 interactions. Pass condition: the same five intents account for at least 60% of work and can be systematised. Stop condition: manual judgement remains unique for every case or errors create unacceptable harm.

Step 3: choose an anchor with a distribution surface

Narita mattered for more than its logo. The airport had stressed information desks, multilingual demand, and a digital path already used by travellers. Tsunagawa says passengers could reach Bebot from the free Wi-Fi landing page. The account supplied three assets at once: buyer credibility, dense real usage, and built-in access.

Score anchor accounts from zero to two on five fields: pain density, user traffic, evidence quality, workflow access, and reference portability. Pursue accounts scoring at least eight out of ten. A famous company with no usage surface scores lower than a niche institution that can expose the workflow daily and explain the result publicly.

Output: a one-page anchor brief naming the buyer, end users, access point, success event, evidence permission, and next-account hypothesis. Pass condition: the pilot creates proof that three named follow-on accounts will recognise. Stop condition: the logo requires bespoke work but blocks usage data, reference rights, and partner learning.

Step 4: embed access before buying traffic

Distribution often fails between purchase and use. A chatbot hidden behind an app-store install would have asked a tired traveller to take a second action. Narita's access path put assistance where a question occurred. That turns an institution's existing traffic into product usage.

Draw the user's five actions before the problem. Insert the product at the earliest step where intent is visible but the workaround has not started. Test QR, browser, inbox, identity-provider, CRM, marketplace, or partner-portal access according to the workflow. Avoid a new account and download unless they are required for security.

Metric: activation from eligible users, not total site traffic. Pass condition: at least 30% of eligible users reach the first value event without staff instruction. Stop condition: the integration adds more steps than the old workaround or the institution cannot identify eligible users.

Step 5: convert institutional proof into partner distribution

In the 2018 interview, Tsunagawa said JR East helped with customer acquisition while Narita shared passenger insights. Those are different partner jobs: one transfers demand and trust; the other improves the product evidence. Do not collapse them into a vague “partnerships” row.

Partner job	Required asset	Measure
Access	Embedded placement in an existing user path	Eligible activations
Trust	Named reference and operational proof	Introductions accepted
Learning	Questions, outcomes, and exceptions	Coverage and correction rate
Acquisition	Three named follow-on accounts	Qualified opportunities

Pass condition: a partner supplies one measurable job and a named owner. Stop condition: the relationship offers only a logo, event appearance, or press release. Bespoke also supplies a useful failure: it discontinued Facebook and WeChat distribution to regain delivery flexibility. A borrowed channel is valuable only while it improves access without controlling the product.

Step 6: expand through adjacent work, not adjacent buyers

BeTrained emerged from the same constraint as Bebot: people could not transfer needed information across language and labour gaps. The workflow records experts with smart devices, transforms their work into multilingual lessons, and tests understanding. A TRY ANGLE EHIME implementation report says the product entered two companies in fiscal 2024. At BEMAC, Bespoke created four safety, one technical, and seven lifestyle lessons in Japanese, English, and Vietnamese.

That field access then exposed inspection work. The founder describes tasks such as scaffold-joint checks and coating-thickness measurement, including a ship where coating inspection can consume more than 1,000 human hours. The current company timeline dates robotics development to March 2026. Treat this as a discovery stage, not proof of commercial scale.

Keep an adjacency queue with four columns: repeated request, shared input, shared buyer, and paid experiment. Advance an item only when at least three customers request it, 60% of the existing data or workflow carries over, and one customer funds a bounded test. Stop when adjacency is only thematic. “It uses AI” is not shared distribution.

What failed, and what remains unproven

• Market timing: investors rejected conversational AI in 2015. Manual tests reduced product uncertainty, not financing risk.
• Borrowed channels: Facebook and WeChat were discontinued because Bespoke wanted more flexibility.
• Reliability: a Christmas update stopped the chatbot for one week and forced founder fulfilment.
• Expansion risk: three wedges share a labour thesis, but chat, training, and robotics have different delivery economics.
• Evidence gap: no public source gives ARR, CAC, payback, gross margin, retention, or revenue by channel.

Your 7-day implementation plan

1. Day 1: choose one observable workflow and recruit 12 recent users.
2. Day 2: log the last action, workaround, delay, consequence, and exact language.
3. Day 3: define a safe manual service boundary and five tracked intents.
4. Day 4: list 20 anchor accounts and score pain, traffic, proof, access, and portability.
5. Day 5: draft one pilot with an embedded-access path and evidence permission.
6. Day 6: identify three trust-transfer partners and assign one measurable job to each.
7. Day 7: review the adjacency queue; advance only requests repeated by three customers.

The week's output is not a campaign volume target. It is one complete chain from observed work to a credible institutional pilot. If any link is missing, fix the link before adding more outreach.

Lead Scorer implementation

Lead Scorer can reproduce the research, qualification, and draft-outreach layer of this motion. It cannot replace field observation, approve a public-sector pilot, disclose a human fulfilment layer for you, or activate a campaign without review.

Start with the ICP and offer context skill. Define the workflow, institution type, end user, compliance boundary, required access surface, and disqualifiers. Create separate lists for anchor accounts, access partners, and trust-transfer partners; combining them destroys the role-specific score.

Use company discovery and web research to build a narrow set. Apply the ICP scoring rubric before paid enrichment. A useful threshold is eight out of ten on the anchor score plus a minimum ICP score of seven. Enrich only keepers, and run contact discovery last for the operator who owns the exact workflow. Credit estimates and human confirmation remain gates.

Then use the signal research dossier and cold email first touch skills to produce one factual message per person. Create only a draft campaign. Review the evidence, role, ask, and disclosure before activation. Replies should be triaged into access, trust, learning, objection, and adjacent-work signals; objections become new research or useful Content Studio briefs rather than automatic follow-ups.

Find 20 airports, stations, municipalities, or infrastructure operators where multilingual support is a visible operational bottleneck. Keep only organisations with an existing digital access surface and a named workflow owner. Score before enrichment. Draft a request for a 30-minute workflow audit, not a product demo. Do not activate or send.

Pass condition: every account has a documented problem, access surface, evidence source, owner, and next-account hypothesis. Stop condition: more than 30% of the list depends on inferred pain or generic innovation titles. The system earns trust through specificity before it spends contact credits.

Saveable checklist

• Recruit people next to the work, not from a generic persona panel.
• Bound manual fulfilment and disclose material human involvement.
• Log intents, exceptions, outcomes, corrections, and repeat use.
• Choose an anchor for traffic, evidence, access, and portable trust.
• Embed the product in an existing user path before buying traffic.
• Give every partner one named distribution job and one metric.
• Advance adjacent products only after repeated paid customer demand.
• Separate company-reported milestones from independently checked facts.

Sources and limits

The main source is the complete This Week in Startups E2320 interview with Akemi Tsunagawa, audited from character zero to its effective end. The Bespoke timeline, company news archive, Narita Airport, the Ehime implementation report, and historical interviews supply web checks.

One contradiction remains: Tsunagawa says the team has 65 people, while the current company page lists 52. The figures may use different dates or definitions, so this course omits headcount. The company timeline says Narita arrived in September 2017, while the contemporaneous release gives a November service date; the course uses only the year. Financial and channel economics remain private. Copy the decision system, not the unsupported scale claims.