The company buys AI licences. Someone runs a lunch-and-learn. A few employees produce impressive examples. Leadership asks everyone to find ways to use it.
Three months later, proposals still take too long, the same manager still chases every handoff, nobody can say whether the investment helped, and the subscription renews anyway.
The leaders wanted this to work. Clients, staffing problems and other demands took their attention. The rollout got a budget, but nobody had the time and authority to change how work gets done.
That gap can explain several problems that later look like poor tools or reluctant employees: decisions wait, learning becomes extra work, and reported errors never get fixed.
AI adds a particular complication to an already difficult change: it can make unfinished work look finished. A polished proposal can contain a guessed price. A confident summary can leave out a condition that changes the decision. The person checking it now has to find what is missing as well as what is wrong.
A demo won't show you whether the AI has the information the work needs, which decisions should stay with people, or whether checking its work costs more time than it saves.
Below are seven places a rollout tends to stall, then a first plan sized for a business like yours.
Seven places a rollout can stall
1. Leaders hand off the learning and keep the decisions
An enthusiastic employee becomes the unofficial AI lead. They can try tools, but they cannot change the process, secure access or settle disagreements between departments.
That employee does extra work, while the leaders who approve changes have barely used the tools themselves.
Your IT provider can handle logins, security and connecting systems. Changing responsibilities and working methods also needs business authority and time. One person can do both, as long as they have the time and authority for each.
2. Nobody defines what the investment should improve
“We need to use AI” can get a budget approved. It gives employees very little direction.
A team can generate more content, summarize more meetings and send more messages without winning another client or removing a bottleneck. Everyone is busy experimenting, but nobody can say what result the effort should be judged against.
3. Staff get new tools without support or influence
A workshop shows what is possible. The next deadline sends people back to the familiar method. Learning to give AI the right background, question its answers and fix its mistakes takes practice. Add it to a full workload, and postponing it makes sense.
A proposal writer catches the same pricing error every week. They are accountable for the proposal, but cannot change the setup and do not know who will fix it. Going back to the old method is a reasonable response.
Running the old method and the new one side by side, with no end date, adds work. If employees expect every improvement to mean a higher quota or a threat to their job, they have reason to keep discoveries to themselves.
4. The AI gets the assignment without the information
Ask for a proposal without current pricing, scope rules or a clear example of acceptable work, and AI may fill the gaps with plausible guesses.
Connecting more files can introduce more ambiguity: two rate cards, an outdated service description, a past proposal with a special discount. A well-written answer can still come from the wrong rate card.
Access needs clear limits: reading a client record, changing it and messaging the client are different permissions. Check with IT or your provider what client information the chosen AI service is permitted to receive.
5. One step gets faster and the business stays slow
AI drafts a proposal in minutes. Then it waits four days for approval. Someone copies the approved details into the project system, where delivery discovers that the scope is incomplete.
The drafting step improved. The customer's wait barely changed.
The real delay may be the approval queue, missing discovery information or a handoff that creates rework. Faster writing leaves those delays intact.
6. Activity gets reported as value
More logins and prompts tell you people are using the tools. They do not tell you whether the business is better off.
Usage numbers also miss what staff keep quiet. In Microsoft's 2024 survey of 31,000 knowledge workers, 52% of people using AI at work were reluctant to admit using it for their most important tasks, and 78% brought their own AI tools. So your licence reports can show low use while staff quietly use other tools, possibly with client information, where you cannot see it.
Even “hours saved” can hide the cost of checking, correction or work transferred to someone else. For example, saving 45 minutes of drafting and adding 40 minutes of checking leaves five minutes, before setup and support costs.
Time saved is not automatically cash saved, either. The benefit depends on what the team does with the hours it gets back. Our article on freed time covers where those hours can go.
7. The demo becomes the expectation
A demo proves something can work with clean inputs. Everyday work brings incomplete notes, unusual clients and exceptions the demo never encountered.
Expect instant, unattended perfection and an ordinary failure can end the experiment. Leave nobody responsible for repairs and even a useful setup can fall out of use as prices, services and tools change.
None of these seven comes from a lack of industry knowledge. Your people know the work. Each one is about people and process: who decides, what should improve, and whether staff get the time and support to change how they work. That matches what large companies report. In BCG's 2024 survey of 1,000 senior executives, about 70% of the problems companies hit with AI came from people and process, 20% from technology and 10% from the AI models themselves. No study we know of ranks these seven by how often they happen; the list is our judgment from rollout work.
What companies seeing returns from AI do differently
In KPMG's Q3 2026 survey of 2,131 senior leaders at companies with at least US$50 million in revenue (US$100 million in Canada), 86% of companies already getting a return from AI had a formal set of tools, rules and checks for how teams use it, shared across teams or the whole company, compared with 31% of companies still experimenting. Those companies were also more likely to check AI's benefit against its cost in every part of the business: 48%, compared with 12% of all respondents. Leaders reported these practices themselves. The survey shows they go together with returns; it does not show they caused them, and the survey did not include small businesses.
OECD research covering more than 5,000 firms with fewer than 250 employees provides a closer view of smaller businesses. In its late-2024 survey across seven countries, including Canada, half of non-users cited a lack of employee skills. Among users, 65% reported improved employee performance. The study did not measure how big that benefit was or whether it paid for itself.
For a smaller business, our recommendation is to put one person in charge with the time to do it, have the staff who do the work help shape it, and track whether it becomes reliable enough to keep. Here is what that could look like.
A workable first plan
Consider a fictional 80-person consultancy that wants qualified prospects to receive proposals sooner without weakening scope or margin.
The leader makes a few decisions
The sales leader chooses the result worth improving, agrees what success would look like, and keeps final pricing and delivery commitments with people. They use AI on one real proposal first, so they know what they are asking the team to do.
They name one person, the implementation owner, to run the change day to day, with authority to coordinate the proposal writers, delivery and IT provider, and remove or defer other work so they have time to do it. If you have no spare person, make an existing manager the implementation owner, block out set hours for it each week, and bring in outside help for the first setup. The leader resolves decisions the owner cannot make and sets a regular time to review progress. They do not need to configure every tool or inspect every draft.
The implementation owner sets up the repeatable work
The owner traces the proposal's full path with the people who write and deliver it, including where it waits and how they handle exceptions. A clearer approval rule or a template may solve part of the problem more simply than AI.
For the parts AI can help with, the owner gives it one approved set of service descriptions, the current price list and a few good past proposals. They arrange the necessary access with the IT provider. When the notes from a client discovery meeting are approved, AI drafts the proposal and the handoff note for delivery, and lists any decision it could not make.
Suppose it finds two conflicting rates. AI should flag the conflict and leave the price blank for an authorized person to set. A human confirms scope, pricing and commitments before the proposal reaches the prospect.
The next proposal uses the same setup, so nobody starts from scratch.
The team helps shape the method
Proposal writers practise, with help, on proposals they already need to finish. They test awkward cases, report problems to the implementation owner and see what changed. If that pricing error returns, a named person investigates it.
The leader explains how the time saved will be used and talks through any expected changes to roles. Staff should not have to guess whether helping means volunteering for more work.
Give AI more work only after it proves itself
Begin with AI preparing drafts and people checking them, and keep the old way ready in case it fails. Compare real proposals and exceptions, then give AI more responsibility for suitable steps only after it has repeatedly met the agreed standard.
For example, checking every heading and formatting choice may become unnecessary. Scope, pricing and commitments still need human approval. Check hardest where a mistake costs money or a client. Drop checks that keep finding nothing.
If checking takes as long as doing the work, do not roll it out further. Improve it or choose a different task. Name who keeps the price list and examples current and who fixes problems, and write down enough for someone else to take over.
The next investment depends on the result
Compare completed proposals of similar complexity before and after the change. Include checking, corrections and the customer's full wait, plus software, setup, training and support costs.
A busy team may learn enough in a month; a low-volume business may need longer. Choose the period around the work.
Expand when the agreed improvement is reached, quality holds and the ongoing cost makes sense. A serious pricing or scope error needs investigation before expansion. Adjust or stop if checking eats the time saved, or if the real delay is somewhere else.
How these roles match the training guide
This plan covers one workflow in one team. Our guides on rolling AI out across a company use three names: the leader, who goes first and owns the program; AI champions, who show their teammates how AI handles everyday work; and a workflow owner for each shared workflow. The implementation owner here is the workflow owner for proposals, and in a company this size is often the first champion too. To spread the method beyond one team, add champions, roughly one for every 15 to 20 staff, as set out in How to train your employees on AI.
Five questions before you invest
Bring these to the person proposing the rollout:
- What business result should improve, beyond using AI more often?
- Who can make the necessary decisions, who has time to implement them, and how can staff get problems fixed?
- What information does AI need, and what happens when it is missing or conflicts?
- Which decisions stay with people, and who maintains the setup after launch?
- How will we judge the complete result, including checking, cost and the customer's wait?
What would this look like in your business?
If you’re planning an AI rollout, or already paying for tools that haven’t changed how work gets done, book a free 30-minute call with Chase.
We’ll talk through what you want AI to improve, where you are today, and whether Sidekick Orchestration is the right help. You don’t need a technical brief.
Still working out where to start? Try the free AI Opportunity Map, a self-assessment that takes about 10 minutes. Your map shows where AI fits your work, how to handle your data and what to do first. Choose “For my organization” and it also includes a rollout plan, a one-page AI-use policy, questions for your insurers and a short brief for your leadership team.
For examples of work AI handles well, see What AI Can Do. To assess a particular task, read Should AI Do This Work?.
