"The most expensive AI project is the one that takes eighteen months to prove it doesn't work." — BlokAI, July 2026
Every discovery call we take in Tokyo eventually reaches the same question, usually asked carefully, halfway through: what is this actually going to cost us? It's a fair question, and most consulting firms answer it badly — either with a shrug ("it depends") or with a six-figure enterprise quote that assumes you're a bank. So let's do the thing nobody in AI consulting seems willing to do in public: put real numbers on the table.
The three tiers, and what actually drives the price
For a small or mid-sized business in Japan in 2026, AI implementation falls into three honest bands.
A scoped automation (¥500K–¥2M). One workflow, one clear owner, one measurable outcome — a voice agent that answers and books your inbound calls, a document pipeline that reads invoices, a support assistant trained on your own knowledge base. This is where most SMEs should start, and where the ROI is easiest to prove. Timeline: two to six weeks. A production system (¥3M–¥8M). Several workflows stitched into your real operations, with the unglamorous parts done properly: data plumbing, permissions, monitoring, a human-in-the-loop path for the 5% of cases the model shouldn't decide alone. This is the ¥5M, 90-day band we build in most often — big enough to change how a team works, small enough to ship before the market moves. Everything above ¥10M is either a genuinely large integration or, far more often, an agency billing you for its own inefficiency. The single biggest cost driver in AI is not model fees or engineering hours. It's scope that was never nailed down — the project that grows a new requirement every fortnight because no one agreed at the start what "done" looks like.Where the money quietly leaks
The visible costs are the least of your worries. The expensive line items are the invisible ones:
- Discovery theatre. Weeks of workshops and slide decks that produce a strategy you already knew. If you're paying for thinking and getting no working software, you're funding someone's overhead.
- Bespoke everything. In 2026 the model layer is close to a commodity. A firm building you a custom LLM from scratch when a well-prompted frontier model would do is selling you complexity, not value.
- The pilot that never ends. "Let's run a three-month proof of concept" is where budgets go to die. A proof of concept that can't show signal in three weeks is telling you something — listen to it.
- Ongoing tax. Ask any quote what month four looks like. Hosting, monitoring, and iteration are real and recurring; a price that pretends they don't exist is a price that will surprise you.
None of this is exotic. It's the same failure mode as any software project — cost tracks clarity. The clearer the problem, the cheaper and faster the answer.
How to buy AI without getting fleeced
A few rules we'd give a friend, even if they hired someone else:
- Buy an outcome, not hours. Insist the contract names a metric — calls answered, hours saved, error rate cut — and a date. Vague scope is the tell.
- Start with one workflow that hurts. Pick the process that costs you the most in time or errors today. One shipped, working system beats a grand roadmap nobody ports to production.
- Demand to see it running before you scale. Ship something small into real use in weeks, measure it, then decide what's next. Momentum is cheaper than planning.
- Prefer boring, proven building blocks. Frontier models, standard infrastructure, off-the-shelf where it fits. Novelty is a cost, not a feature.
Done this way, most Tokyo SMEs get a working, measurable AI system live for a fraction of what a Big-Four-style engagement quotes — and they get it this quarter, not next year. The point of implementation isn't to look sophisticated. It's to have something running by month's end that pays for itself, then to build the next thing on top of proof instead of hope.
That's the whole philosophy: ship fast, measure honestly, scale only what works. It's also, not coincidentally, the cheapest way to do it.
Want to implement AI that actually delivers? Start at blok-ai.com/free-discovery