task-ops
Every project tool now has an AI button, and most of the writing about it answers a question nobody running a small team is asking. The question is not whether AI will change the profession. It is narrower and more immediate: of the twenty or so recurring chores that fill a coordinator's week, which ones can be handed over this month, and what does handing them over cost.
That question has an answer, because the features have stopped being demos. They ship, they are priced, and the pricing models differ enough that two teams paying the same headline seat price can end up with wildly different AI bills. What follows is the shape of the capability, the chores it genuinely absorbs, the ones it does not touch, and a test that produces a decision rather than an impression.
Strip the marketing off the AI sections of the major tools and four distinct capabilities remain. They are worth separating, because a team can get real value from one and none from the other three.
Text generation. Writing a status summary from the state of a board, drafting a task description from a one line title, turning a long comment thread into three bullet points. This is the most reliable category and the least interesting, because the output still needs reading before it goes anywhere.
Extraction. Taking unstructured input, a meeting transcript, a customer email, a page of notes, and producing structured items: task titles, assignees, due dates. This is where the visible time saving lives for most coordinators.
Question answering over the workspace. Asking what is blocked, which items slipped this week, what a particular client has open. Instead of opening five boards and filtering each one, the question is typed once.
Scheduling and routing. Auto assigning by workload, reordering a personal queue by deadline and estimate, auto scheduling tasks into calendar gaps. Some tools treat this as the whole product rather than a feature.
A team that mainly needs the third capability should not pay for a tool built around the fourth, and vice versa. Deciding which of the four matters is the first useful step, and it can be done before any trial starts.
Four recurring jobs come off the plate quickly and stay off.
The weekly status write up. A coordinator who assembles a Monday summary by reading boards and rewriting card titles into prose is doing work an extraction and generation model does acceptably. The draft still needs correcting, but the blank page problem disappears, and a task that took forty minutes takes ten.
Turning notes into items. After a planning call there are twelve things to do buried in a page of notes. Typing them into a board one at a time is the single most mechanical task in project coordination. Pasting the notes and getting twelve draft cards, then deleting the three that are wrong, is faster by a wide margin.
Answering where things stand. Most interruptions a coordinator absorbs are lookups. A colleague asks whether the translation came back, whether the invoice went out, whether the client approved the layout. Being able to ask the workspace and get an answer moves those interruptions off the coordinator entirely, because the colleague can ask it directly.
Finding what went quiet. Items that stop moving are the thing boards are worst at surfacing. A card sits in a column for three weeks and nobody notices until someone downstream is blocked. Asking which items have had no activity since a given date is a query, and a model is good at running it and summarising the result.
Notice what these four have in common. They are all reading, restating and retrieving. None of them requires a decision.
The limits are consistent across tools, and they are not going to be solved by a better model, because they are not model problems.
Estimation. A model can produce a duration for a task. It has no access to how long that task took the last four times, unless the team recorded it, and teams that record it accurately do not need the estimate. The number that comes back reads confidently and carries no information.
Priority when people disagree. Two stakeholders both want their work first. This is not an information problem to be resolved by better ranking. Somebody has to decide and then tell the loser. No tool removes that conversation.
Accountability. An AI generated status report that turns out to be wrong is the coordinator's error, not the tool's. This changes the review burden rather than eliminating it. Output that goes to a client needs reading line by line, which caps the saving on client facing work well below the saving on internal work.
Anything the board does not know. This is the largest limitation and the easiest to overlook. If half the work lives in chat messages and email threads, the model sees half the project, and it does not announce that it is guessing. A team with a partly maintained board gets confidently wrong answers, which is worse than no answers.
The headline seat price no longer tells the whole story, because vendors have settled on four different ways to charge for AI. The figures below were taken from the vendors' own pricing pages in September 2026, in the currency each page showed.
| Tool | How AI is charged | Published figures |
|---|---|---|
| ClickUp | Separate per seat add on | Brain at $9 per user per month, or $7.20 on annual billing. A wider Everything AI tier at $28 per user per month, or $22.40 annual. Base plans: Unlimited $7 annual and $10 monthly, Business $12 annual and $19 monthly |
| Asana | Metered per request | Starter and Advanced include AI Teammates and Asana Dash at 5 requests per user per month, capped at 50 per account. Further requests cost $0.50 each billed upfront or $0.60 each billed monthly. AI Studio comes with 50,000 credits per account per month on Starter, 75,000 on Advanced |
| monday.com | Credits by tier, from Basic upward | 1,000 AI credits a month on Basic, 2,000 on Standard, 3,000 on Pro. The free plan is capped at 2 seats and 3 boards and has no AI |
| Notion | Bundled into the plan | Free and Plus carry a trial of Notion AI. Business adds AI Meeting Notes and Enterprise Search. Agents are free to try, then $10 per 1,000 monthly Notion credits |
| Motion | AI is the product | Pro AI at $19 per seat monthly or $12.73 annual with 7,500 credits a month. Business AI at $29 monthly or $19.43 annual with 15,000 credits. No free plan, trial only |
Two consequences follow. The first is that metered and credit based pricing makes the monthly cost a function of usage, which nobody can forecast before the trial. A team that finds the feature useful is a team whose bill goes up. The second is that a per seat add on is charged for everybody on the plan, including the five people who will never press the button, so the effective cost per actual user can be several times the sticker price on a small team.
Asana's free tier is capped at 2 users and monday.com's at 2 seats, which means for most teams the AI question and the paid plan question arrive together rather than separately. Worth checking the base pricing of whatever is under consideration before the AI line item, because the add on is usually the smaller of the two numbers.
Every capability above depends on the board being an accurate picture of the work. This is a harder condition than it sounds, and it is the reason most AI trials in project tools end inconclusively.
A board is accurate when three things hold. Every piece of committed work exists as an item, including the small things that currently live in somebody's head. Items carry an owner and a date, not just a title. And status is updated when the work moves, not retroactively before a review.
Teams that already meet those conditions get immediate value from AI features, which is why the case studies always feature them. Teams that do not will get answers assembled from a partial record. The fix is not a better tool, and it is not a rule that people should be more diligent. It is usually that the work is spread across a board, a chat app and a spreadsheet, so keeping any one of them current is thankless. Consolidating first, then adding AI, produces a very different result from doing it the other way round. Looking at what a single board needs to hold is a more useful exercise than comparing AI feature lists.
There is a second route that the vendor pricing tables do not cover. Rather than paying for AI inside the project tool, the assistant a team already pays for can be connected to the board directly.
The mechanism is Model Context Protocol, an open standard for letting an assistant call out to an external system. A project tool that publishes an MCP endpoint can be added to ChatGPT or Claude as a connector, after which the assistant can read boards, create cards from a pasted set of notes, move items between lists, edit fields and read scheduled tasks with their dates. On the ChatGPT side this currently runs through developer mode, which is available on paid plans. In Claude it is added in connector settings, and Claude Code adds it from the command line.
The practical difference is where the cost sits. A team with existing assistant subscriptions pays nothing additional per seat in the project tool, and the assistant keeps the rest of its context, meaning the same conversation that drafted the client email can create the follow up cards. Permissions carry over from the project tool, so boards a person cannot see stay invisible to the assistant, and billing and token settings stay out of reach. For teams whose real complaint is copying between an assistant and a board, connecting the two removes the actual friction, and no AI add on is involved.
On the other side of the ledger, this route gives no auto scheduling and no workload balancing. Those live inside the tools that are built around them. It is a good answer for extraction and question answering, and not an answer for routing.
Impressions from a trial are worthless because everything is impressive for three days. A test that answers the question looks like this.
Pick the single chore that costs the most hours. Not a list, one chore. For most coordinators it is the status write up or the notes to cards conversion. Time it honestly for one week the current way, including the interruptions.
Then run the same chore through the AI route for one week, and record two numbers: minutes spent, and corrections needed. The second number is the one that decides it. A tool that saves thirty minutes and introduces one error a week into a client report has not saved anything.
At the end, compare against the actual monthly cost including metered usage, not the advertised seat price. If the saving is under an hour a week per person, the honest answer is that the board needs consolidating before AI is worth paying for.
Before comparing AI features, check whether one board holds all of the committed work, because every capability in this article reads from that board and nothing else. If work is currently split across a board, a chat thread and a spreadsheet, fix that first and the AI question gets much easier to answer. Pinateca keeps kanban, Gantt, calendar and chat on one board and connects to ChatGPT and Claude as a connector, so a team can test the extraction and lookup chores without adding a per seat AI charge.
It depends on which of the four capabilities the team needs. Extraction, turning notes into tasks, and workspace lookups pay off at any size, because both chores scale with the number of meetings rather than the number of people. Auto scheduling and workload balancing pay off when there are enough people and enough tasks for the allocation to be genuinely hard, which is usually not the case at five.
Turning unstructured notes into structured tasks. It is the most mechanical part of coordination, it happens after every meeting, and a wrong output is cheap to fix because the card is deleted in one click. Status summaries save similar time but need closer review, especially anything that reaches a client.
Not on the evidence of what the current features do. Everything shipping today is reading, restating and retrieving. Estimation without historical data, deciding between two stakeholders who both want priority, and carrying responsibility for a wrong report are all still human work, and they are the parts that take the judgment.
Partly, and less than the headline suggests. Notion carries a trial of Notion AI on its free plan, and ClickUp's free tier offers trial access to Brain. monday.com states that AI is available from the Basic plan upward, and its free plan is capped at 2 seats and 3 boards. Asana's free Personal plan is capped at 2 users, so for most teams the paid plan decision arrives before the AI decision.
Yes, where the project tool publishes a Model Context Protocol endpoint. The tool is added as a connector, after which the assistant can read boards and create, move and edit cards under the same permissions the person already has. This avoids paying a second time for AI inside the project tool, though it provides no auto scheduling.