Where Is the Excel Solver, and When Does AI Do the Job Faster

Short answer: Data tab, Analysis group, all the way on the right. If it is not there, it is because Solver ships as an add-in that is off by default, not because anything is broken.

Turning it on

1. 01File → Options → Add-ins.
2. 02At the bottom, next to "Manage", choose "Excel Add-ins" and click Go.
3. 03Tick "Solver Add-in" and click OK.
4. 04Go back to the Data tab. Solver now sits in the Analysis group, usually next to Data Analysis.

That is the whole install. No download, no admin rights — it ships with every desktop copy of Excel, just switched off until you ask for it.

What Solver actually does

Solver finds the values for a set of input cells that push one target cell to a maximum, a minimum, or a specific number — subject to constraints you define. Say you buy the same part from three suppliers, each with a different price and a monthly capacity limit, and you need 10,000 units at the lowest total cost. Solver is built exactly for that: minimise the total-cost cell, let the three order-quantity cells change, and constrain each to its supplier's capacity and to a combined total of 10,000.

That is a linear program — the classic use case, and the one Solver's default Simplex LP method is built to solve reliably. It also handles nonlinear problems, but those can settle on a local optimum rather than the true best answer, which matters later. Excel ships three solving methods in the dropdown at the top of the dialog: Simplex LP for straightforward linear problems, GRG Nonlinear for smooth nonlinear ones, and Evolutionary for anything with a messier, non-smooth relationship between the changing cells and the objective. Picking the wrong one is the most common reason Solver reports "no solution found" on a problem that clearly has one.

Where AI actually fits in

The optimisation itself still happens inside Solver's algorithm — that part is not a job for a language model, and asking one to "just solve it" skips the part of the tool that is actually reliable. Where AI genuinely helps is the setup: turning a plain-language problem into the objective cell, the changing cells, and the constraint list Solver needs, which is the step most people find fiddly the first few times.

That is a prompting task, and it follows the same rules as any other prompt: be specific about the job, not vague about the outcome. OpenAI and Anthropic both frame good prompting the same way in their own guidance — state the goal, the inputs, and the shape you want the answer in, rather than a one-line request and hoping. How to write a prompt that works on the first try covers the general version of this.

A worked example

Weak: "Solve this in Excel."

Better: "I have three suppliers with per-unit costs of $4.20, $3.90, and $4.60, and monthly capacity limits of 5,000, 4,000, and 6,000 units. I need exactly 10,000 units total at the lowest combined cost. Walk me through setting this up in Excel Solver: which cell should be the objective, which three cells should be the changing cells, and what the constraint list should say."

The second version gives the model everything it needs to describe an actual Solver setup rather than a generic explanation of what Solver is. You still open Solver yourself, still enter the objective cell, changing cells, and constraints, and still click Solve — the model saved you the trial and error of getting that dialog right the first time.

This kind of repetitive setup work — the same three-field dialog, filled in slightly differently every month — is exactly the category find the repetitive part is about: work worth automating once you notice the pattern, rather than re-deriving from scratch each time.

Where this goes wrong

Two separate failure points, and they are easy to conflate.

• —Solver itself can settle on a local optimum for nonlinear problems, or report "no feasible solution" when the constraints actually contradict each other and need loosening, not more solving.
• —An AI model asked to check its own suggested setup will often just agree that it looks right, which is not the same as it being right. Anthropic has measured this tendency and named it sycophancy — models trained on human approval tend to produce approval-shaped answers, including when you ask them to grade their own work.

The same caution applies here as anywhere else an AI tool hands you something that looks finished: what AI is actually bad at covers this pattern in more detail, and it does not go away just because the output this time is a spreadsheet formula instead of a paragraph.

Checks before you trust the result

1. 01Run a known-answer test case first — a simplified version of the problem where you can work out the right answer by hand, and confirm Solver lands on it.
2. 02Check the Solver Answer Report: are the constraints you expected to bind actually binding, or is the model quietly not using one at all?
3. 03Confirm the units match across every cell. A capacity in "units per month" mixed with a cost in "per case" produces a confident, completely wrong optimum.
4. 04Before you even open Solver, read what your objective and constraint cells actually contain in the formula bar — especially if an AI tool helped draft them. A reference one row off is invisible in the cell itself and obvious the moment you look there.
5. 05If the objective value looks suspiciously good, that is the moment to check it by hand, not the moment to trust it. How to check an AI answer when you are not the expert is the general version of this instinct, and it applies here too — the source of the number was a tool, not a colleague, but the habit of checking is the same.

Solver can also generate a Sensitivity Report alongside the Answer Report — tick it in the results dialog before you click OK. It shows how much each constraint is actually costing you, in the units of your objective, which is usually the more useful number for a real decision. A supplier constraint with a shadow price near zero is not actually limiting you; one with a high shadow price is worth renegotiating before you touch anything else in the model.

Solver tells you the best answer to the model you built. It cannot tell you whether you built the right model.

This is also why how accountants and financial analysts actually use AI at work treats it as a drafting and setup tool rather than a source of final numbers — the same discipline shows up in how to use AI as a financial analyst, and it holds for a spreadsheet optimisation as much as for a forecast. AI is reshaping a lot of this repetitive, structured spreadsheet work, which Stanford Graduate School of Business has written about directly — the boring, mechanical setup shrinks; the judgement about whether the model is actually right does not.

What to do Monday

Find one decision you are currently making by trial and error in a spreadsheet — a staffing schedule, a budget split across projects, an order quantity across suppliers — and set it up properly in Solver instead, using an AI tool to help you draft the objective and constraints from a plain description first. Businesses making automated decisions at any scale are also expected to be able to explain how those decisions were reached — NIST's AI risk management framework is the reference point for that expectation — which is one more reason to be able to show your constraints, not just your answer.

Coursium teaches this kind of practical, tool-specific skill directly — describing a real problem clearly enough that a tool can actually help with it. Stay ahead of AI by learning the tools on your phone.