Stop bending your operations around shiny tools. Start with the problem, then build the right solution.
There is a very simple reason many AI projects fail. The technology arrives before the problem has been properly understood.
Someone sees a demo. A board member reads a headline. A competitor says they are “using AI”. Suddenly, the business is trying to wedge artificial intelligence into places where the real issue might be a broken workflow, poor data, duplicated admin, slow approvals, clunky software, or a process that should have been fixed three years ago.
At Coal Face AI, we take the opposite approach. We do not start by asking, “Where can we put AI?” We start by asking, “Where is the business leaking time, money, quality, customer trust, or opportunity?” Only then do we decide whether AI is the right answer.
Sometimes AI is part of the answer and sometimes AI is not the answer at all.
Not sure where AI fits in your business? Start with a proper operational review. Coal Face AI can help you identify the workflows, bottlenecks and pain points where technology could actually make a measurable difference.
The problem with “AI-first” thinking
“AI-first” sounds bold. It sounds modern. It sounds like the sort of phrase that belongs on a glossy slide with a rocket icon and a stock image of a robot looking thoughtful. But in practice, AI-first can quickly become business-second.
That is where things start to go wrong. The tool becomes the centre of gravity. The business is expected to adapt around it. Staff are told to use it. Workflows are changed to accommodate it. Data is dragged into places it does not belong. Nobody is entirely sure what success looks like, but everyone agrees the organisation is “doing AI”.
The research is increasingly clear on this. Gartner reported that at least 50% of generative AI projects had been abandoned after proof of concept by the end of last year, citing poor data quality, inadequate risk controls, escalating costs and unclear business value as major reasons.
McKinsey’s 2025 State of AI survey tells a similar story. AI use is spreading quickly, with 88% of respondents reporting regular AI use in at least one business function. But most organisations are still experimenting or piloting, and only 39% report enterprise-level EBIT impact. McKinsey also found that redesigning workflows is a key success factor among higher-performing AI organisations.
Bad AI projects usually start with the wrong question
The wrong question is:
“What AI tool should we buy?”
A better question is:
“Which part of our business is costing us more than it should?”
That might be a customer service team drowning in repeated enquiries. It might be sales leads getting lost between email, CRM and spreadsheets. It might be engineers spending hours writing reports after every site visit. It might be managers manually chasing updates that should be visible in real time. It might be a finance process that relies on six versions of the same spreadsheet and that one person who knows where everything is saved. That is where AI implementation should begin.
Not with a model. Not with a vendor. Not with an app. With the ugly bit of the business that everyone never talks about because “that’s just how we’ve always done it”. Those ugly bits are often where the value is.
If your team has a workflow that everyone complains about but nobody has properly fixed, that is probably a better starting point than buying another AI tool. Coal Face AI can help map it, challenge it, and design a better way forward.
The business should define the AI, not the other way round
A good AI solution should feel like it belongs inside the business. It should reflect the organisation’s customers, data, systems, risk appetite, staff capability, tone of voice, approval routes, compliance needs and commercial goals. It should reduce friction, not add another dashboard. It should make work easier, not force people to become part-time prompt engineers just to get through the day.
That means the design process has to start with context.
For example, a customer-facing AI assistant in a regulated business needs very different guardrails from an internal document summarisation tool. A manufacturing scheduling assistant needs different data foundations from a sales proposal generator. A medical triage tool, a legal research assistant, a stock enquiry agent and a warehouse optimisation system are not just “chatbots with different prompts”.
They are different business systems with different risks.
The Air Canada chatbot case showed this clearly. The airline was ordered to compensate a customer after its chatbot gave inaccurate information about bereavement fares. The tribunal made the point that the chatbot was still part of Air Canada’s website, and the company was responsible for the information it gave.
If an AI system touches customers, decisions, money, policy, safety, contracts, personal data or reputation, it needs to be designed around the business context it operates in. Otherwise, you risk introducing a new failure point.
AI does work when it is matched to the right task
This is not an anti-AI argument. Far from it.
AI can produce serious gains when it is applied to the right work in the right way. A major NBER study of 5,179 customer support agents found that access to a generative AI assistant increased productivity by 14% on average, with a 34% improvement for novice and lower-skilled workers. It also found evidence of improved customer sentiment and employee retention.
That matters because it shows where AI can be powerful: not as a vague productivity fairy, but as a specific tool inside a specific workflow, supporting specific users on specific tasks.
The Harvard and BCG “jagged frontier” research makes the same point from another angle. In tasks where AI was well suited, consultants using GPT-4 improved speed by over 25% and human-rated performance by over 40%. But the research also warned that AI performs unevenly across different types of work, excelling in some areas and falling short in others.
In plain English: AI is not equally good at everything and this is why fit matters.
A good implementation partner does not simply ask whether AI can do something. They ask whether it should, whether it can do it reliably enough, whether the business process around it is ready, and whether the benefits justify the cost, risk and change required.
Where companies go wrong
Most poor AI implementations do not fail in one dramatic explosion. They usually fail subtly in the background.
They fail because nobody defined the problem properly. They fail because the data is scattered, duplicated or untrusted. They fail because the team expected the AI system to fix a process nobody truly understands. They fail because the system was not integrated into daily work. They fail because staff were not trained. They fail because governance was added at the end, like a seatbelt fitted after the crash. They also fail because businesses confuse novelty with value.
McDonald’s ended a trial of AI drive-through ordering across more than 100 US restaurants after testing the technology with IBM. The company did not say AI had no future in drive-throughs, but the trial became associated with wrong-order stories and showed the difficulty of applying automation in noisy, high-pressure, customer-facing environments.
DPD also had to disable part of its AI chatbot after a customer prompted it into swearing and criticising the company. DPD said an error occurred after a system update and that the AI element was disabled while being updated.
These examples are not reasons to avoid AI. They are reasons to implement it properly.
The lesson is not “AI is bad”. The lesson is “AI without operational design, testing, governance and fallback routes is a risk”.
Start with the pain points
Before building anything, a serious AI project should begin with questions like these:
- What work is repetitive, slow or unnecessarily manual?
- Where do customers experience delays?
- Where do staff duplicate effort?
- Where are decisions being made with poor visibility?
- Where does information get lost between people, systems or departments?
- Where is quality inconsistent?
- Where are skilled people spending time on low-value admin?
- Where are risks currently being managed by memory, habit or heroic effort?
That last one is important. Many businesses are held together by informal workarounds. Someone knows which spreadsheet is the “real one”. Someone remembers which customer needs special handling. Someone manually checks the thing the system should have caught. Someone copies data from one platform into another every Friday afternoon because the two systems do not talk to each other.
AI can help, but only if the workflow is properly understood first.
Coal Face AI starts by getting into the reality of how your business works. Not the process map you wish you had. The actual one. The messy one. The one full of workarounds, delays, duplicated admin and “we’ve always done it this way”. That is where better systems begin.
Then decide what sort of solution is actually needed
Once the pain points are clear, the solution might be AI. It might be automation. It might be better software integration. It might be a decision-support dashboard. It might be a custom internal tool. It might be a customer-facing agent. It might be a simple workflow redesign with a small amount of AI doing one useful job in the background.
The business need should shape the solution. A company with poor lead handling probably does not need “an AI strategy”. It may need a better intake process, CRM integration, automated qualification, intelligent follow-up and a clear handover route to sales.
A service business with missed calls may not need a generic chatbot. It may need an AI receptionist that understands services, appointment rules, location constraints, escalation routes and customer intent.
A manufacturing firm with inconsistent reporting may not need a large language model bolted onto SharePoint. It may need structured data capture, standardised reporting, exception detection and a dashboard that helps managers act faster.
A professional services firm may not need everyone using random AI tools. It may need approved workflows, secure document handling, reusable prompt patterns, quality assurance and staff training that reflects the work people actually do.
Fit also means governance, security and trust
A solution that fits the business must also fit its risk environment.
That means looking at data protection, access control, auditability, human oversight, cyber security, model limitations, fallback processes, and who is accountable when something goes wrong.
NIST’s AI Risk Management Framework is designed to help organisations incorporate trustworthiness into the design, development, use and evaluation of AI systems. NIST has also published a generative AI profile to help organisations identify risks specific to generative AI and align risk management actions with their goals and priorities.
The UK government’s AI assurance guidance makes a similar point. It describes AI assurance as a way to measure, evaluate and communicate whether AI systems are trustworthy, including whether they work as intended, what limitations they have, and how risks are being mitigated.
This is where many businesses underestimate the work involved.
A useful AI system is not just a clever front end. It needs the right architecture, the right policies, the right testing, the right human oversight and the right operational controls.
At Coal Face AI, that is part of the implementation conversation from the start. Not because governance is fashionable, but because real businesses need systems they can trust.
The Coal Face AI approach: diagnose, design, deliver
At Coal Face AI, our approach is deliberately practical.
First, we diagnose the business problem. That means looking at workflows, systems, users, customer journeys, data, costs, risks and commercial impact. We want to know where the pain is and why it exists.
Then we design the right solution. That may involve AI agents, workflow automation, bespoke software, decision-support tools, data pipelines, customer interaction platforms, or a mix of technologies. The design is based on the job to be done, not on forcing a fashionable tool into the room.
Then we deliver it properly. That means implementation planning, technical build, testing, integration, staff adoption, governance, iteration and measurement.
The aim is not to “do AI”. The aim is to make the business work better.
If you are under pressure to adopt AI but do not want to waste money on the wrong thing, Coal Face AI can help you move from vague ambition to a practical, costed, deliverable solution.
What a good AI fit looks like
A well-fitted AI solution has a few obvious signs.
- People use it because it helps, not because they were told to.
- It reduces steps instead of adding them.
- It connects to existing systems where sensible.
- It has clear boundaries.
- It knows when to hand over to a human.
- It produces measurable value.
- It improves over time.
- It is secure enough for the environment it operates in.
- It has an owner.
- It can be explained to the people affected by it.
Most importantly, it solves a problem the business already cared about before anyone mentioned AI.
That is the real test.
If the only reason a project exists is “because AI”, it probably needs more thinking.
The boardroom question should change
For the last couple of years, many leadership teams have been asking:
“What is our AI strategy?”
That is not a bad question, but it is often too broad to be useful. It leads to workshops, frameworks, roadmaps and ambition statements. Those things have their place, but they do not automatically fix broken operations.
A sharper question is:
“Which business problem are we solving first, and what is the best way to solve it?”
That question forces discipline.
It brings the conversation back to cost, revenue, risk, quality, speed, customer experience and staff workload. It makes the business case clearer. It makes the technical design better. It makes adoption easier because people can see the point.
And it helps avoid the classic mistake of making the business fit the AI solution.
The AI solution should fit the business. Every time!
Final thought
AI is not a badge you pin to your company to prove you are modern. It is a capability. Used well, it can remove friction, improve decisions, reduce admin, support staff, serve customers faster and create new commercial advantage. Used badly, it becomes another expensive system people quietly work around.
The difference is not usually the model. The difference is the thinking before the model.
Start with the pain. Understand the workflow. Challenge the process. Assess the data. Design around the users. Build around the business. Govern it properly. Measure the outcome.
That is how AI becomes useful.
And that is where Coal Face AI comes in.