Ali Parandeh — a chartered mechanical and software engineer, O’Reilly author (his book covers building generative AI services with FastAPI), previously head of engineering and now running his own AI advisory for heavy industry — interviewed on TwoSetAI (30 min). His market is construction, automotive, energy, aerospace and infrastructure: sectors where a bad model is not a bad decision, it is a physical failure. He came back to the show to talk about how those companies actually buy AI, how to scope projects that cannot be allowed to fail, and the job title his business model ends up needing.

Why heavy industry is two or three years behind

  • IT, finance, retail, advertising and creative adopted fast — they already have the talent, and the downside of a mistake is contained
  • In engineering sectors a mistake has consequences, so the culture and the procurement process are risk-averse by default and they will not move on a demo alone
  • They are under real pressure: they know competitors and adjacent sectors are getting returns, but they lack clarity on what AI can do, what it cannot do, and what happens when it fails
  • The result is a two-to-three-year lag — and a lot of demand for someone who can translate between the two worlds
  • His model is two-pronged: raise awareness through the professional institutions the sector already trusts (Institution of Mechanical Engineers, Engineers Ireland), then consult on use-case discovery, pilots and scaling

The number-one misconception

  • Executives believe they cannot use AI because their data is not safe, not secure, or they do not know how it will be used
  • So most of them stop at Copilot, because M365 is the standard enterprise platform and Copilot is the only approved chatbot — it is the minimum viable AI program, not a strategy
  • An org-wide Copilot rollout differentiates nobody, and most of those employees have no idea how to use it beyond generating some text
  • The value is in narrow, self-built copilots for specific repeatable workflows — weekly and quarterly reporting, the admin that eats 20-30% of an engineer’s time
  • Meanwhile the pro-AI people quietly use their personal ChatGPT or Claude anyway, outside the sanctioned stack

Where the budget actually goes

  • Not into generic assistants, but into systems that unlock value already sitting in the plant. Predictive maintenance is the highest-adoption use case in the sector
  • The data is bespoke: a fiberglass manufacturer using computer vision to predict a line failure seconds in advance; an automotive plant predicting honing-ring failure during gearbox production
  • Bespoke data means a bespoke model. “You can’t just place the data into Copilot and get it to work” — wrong model, wrong shape, and a lot of it. It needs a purpose-built ML model plus the domain expertise to interpret what the data even means
  • Off-the-shelf, consumer-facing tools will not solve it. These companies are sitting on custom data only they understand, and that is exactly why they have to build
  • Product bets on top of that: he is building enterprise requirements-management software for large construction and engineering projects — analyzing requirements, preparing stage-gate packs, and evidencing regulatory compliance

The three blockers

  • Per O’Reilly’s research, and his own experience: use-case identification — you cannot ideate on AI without understanding its capabilities and limitations
  • Data quality — the data may not exist (you install sensors) or may not be in usable shape
  • Talent — most of these firms have never funded a data-science or AI engineering team, and have not upskilled existing engineers
  • The build-or-buy line: sometimes you build in-house because you intend to experiment seriously, sometimes you subcontract to specialists. Usually it is both

The dark room with a lantern

  • Most AI prototypes never reach production. His answer for convincing a CFO: stop arguing about certainty and cap the cost of finding out
  • Engineering firms are waterfall-native — a railway from A to B has standards, known risk, and a stage-gate process that works. AI projects are R&D, and that process does not transfer
  • His metaphor: “With normal engineering projects, sometimes the room is lit so you can plan ahead. With AI, it’s mostly a dark room.”
  • So you walk the room with a lantern. Until you walk it, you do not know where the walls, the table, or the exit are
  • The method is agile over waterfall: pick the team and the time box, then let the scope move inside it — “there are three of us, you have one month, what can we do?”
  • Every sprint has to end in a finding, and negative findings are still deliverables: this direction does not work with this data is money well spent, because it stops the next expensive mistake
  • Allow more spent risk as clarity grows. That is what releases senior leadership budget

Four ways agent projects fail

  • They skipped the discovery phase and rushed straight to productionizing an agent
  • They did not red-team it — edge cases and bad cases were never handled in the pilot
  • They built no adoption strategy. A tool nobody uses is a write-off regardless of quality; some firms mandate usage, others run training and webinars
  • They shipped without a security layer and without production-grade reliability
  • Add ROI and user adoption to the discovery checklist and the failure rate drops — the same red-team-and-measure discipline that shows up in every honest postmortem

How the consulting model scales

  • Stage 1 awareness and training, Stage 2 discovery and use-case mapping (he runs 10-day discovery sessions with manufacturing clients), Stage 3 pilots, Stage 4 build-for-you
  • Not every client needs all four: some arrive having done the awareness work and just want pilots built
  • He starts at low commitment and low risk, then increases both as ROI and clarity appear
  • Consulting does not scale on headcount alone. His plan: software first, to release capacity where pilot demand becomes the bottleneck, then hiring
  • The role that fills the gap is the forward deployed engineer — an AI-fluent builder embedded in the client’s team, shipping and iterating pilots quickly. It is the model a16z-backed firms in these sectors use, and it is where he expects his own hiring to go

Craft is the thing AI should not take

  • “If you can do it without AI, do it without AI.” He hand-codes his own sites and labels them human-made; he wrote his book without AI
  • His prediction: handcrafted artifacts gain value as AI-produced output floods the market — the same instinct behind businesses that advertise you will speak to an actual human
  • He opens every customer-service bot conversation with “talk to an agent” to bypass the whole flow. The purpose of AI is to remove the parts you do not want to spend time on, not the craft itself
  • Identity is part of it: the satisfaction of ownership is joy, pride and pain together. “Without the pain, all the happiness would feel meaningless”
  • It also cuts against his own commercial interest — he could run advisory and training only, and still builds software because that is the part he loves

Fundamentals still matter

  • He still recommends the foundational texts: Chip Huyen’s AI Engineering and Designing Machine Learning Systems, and Hands-On Machine Learning with Scikit-Learn, Keras and PyTorch
  • Why, when models can write the code: AI makes mistakes, and expertise is the only thing that catches them. Craft and expertise get more valuable, not less
  • Career advice he wishes someone had given him at 20: Dan Priestley’s Entrepreneur Revolution, Key Person of Influence, Oversubscribed and 24 Assets — become the person the industry calls when it needs this, and publish so the opportunities come to you
  • Build the business by doing business; that becomes your MBA

With with normal engineering projects, sometimes the room is lit so you can plan ahead. With AI, it’s mostly a dark room. So you have a lantern in your hand and you need to walk the room. Until you’ve walked the room, you have no idea where the walls are, where the table is, where the door is to exit.

His closing note: this is a new information revolution, not a fad to wait out. The people who plan ahead and surf the wave beat the ones who stand still and get flattened by it — with the caveat he repeats all the way through, that he promotes the limitations more than the capabilities.