“Let’s throw AI at it.”
I understand the instinct.
Companies have thousands, sometimes millions, of individual charges flowing through invoices, contracts, portals and financial systems.
There is almost certainly waste hiding in there.
So give all that data to AI and ask:
“Where can we save money?”
And AI will probably find some good ideas.
That’s exactly what worries me.
Not because the recommendations will necessarily be wrong.
They may be excellent.
The danger is that a few good savings recommendations can create confidence that you’ve comprehensively analyzed your spend — when the system may have looked at only a fraction of the ways savings could exist.
A few good savings recommendations don’t prove you have good savings coverage.
After decades of working on cost optimization and human-centered expense management processes — including writing Principles of Telecom Expense Management — I’ve come to believe this distinction is incredibly important as companies begin applying AI to cost reduction.
Before you automate cost optimization, you need to understand cost optimization.
Or, put another way:
“Let’s throw AI at it” works much better when leaders understand what “it” actually involves.
The Savings Coverage Problem
Think about one charge on one invoice.
How many questions would you need to ask before you could confidently say:
“Yes. We should be spending every penny of this.”
- Billing: Was the charge billed correctly and according to contract?
- Usage: Is the service actually being used?
- Validity: Is it still associated with an active employee, device, location, or other legitimate business need?
- Duplication: Are we already paying for another service that provides the same capability?
- Configuration: Is the service appropriately sized and configured for what we actually need?
- Pricing: Is the price competitive?
- Discounts: Are we receiving every available discount?
- Alternatives: Is there a lower-cost plan, commercial structure, or alternative service that would achieve the same business outcome?
And perhaps most importantly:
Do we actually have enough information to answer all those questions?
That’s a very different problem from asking AI to “find some savings.”
I think of it as Savings Coverage:
How do you know your AI and human system is looking for and capturing all the different kinds of savings that could exist?
And there are really two parts to it.
Discovery Coverage
Did we investigate all the places an opportunity could exist?
Realization Coverage
Did we follow every valid opportunity through until the saving was actually captured and verified?
You need both.
Because missing an opportunity produces no saving.
And finding an opportunity but failing to implement it also produces no saving.
Cost optimization is a coverage problem
This is something we learned long before generative AI.
There are lots of ways for money to leak out of an organization.
Some are obvious.
Some aren’t.
You can be paying for something nobody uses.
You can be paying for something nobody needs.
You can be paying for something you already receive somewhere else.
You can be paying the wrong contractual rate.
You can be paying the correct contractual rate — but still paying too much.
You can be missing a discount.
You can have the wrong configuration.
You can have a service associated with a location that closed six months ago.
You can have a service everyone believes was disconnected that the vendor is still billing.
You can have a charge that looks completely legitimate on the invoice, while information sitting in a vendor portal shows that the service has had no usage.
You can have a perfectly accurate invoice and still have a bad deal.
These aren’t variations of one problem.
They are different savings tests.
And that matters when we start designing AI systems to find savings.
AI may find exactly what you ask it to find
Imagine giving an AI system a year’s worth of invoices.
It analyzes them and finds:
- billing anomalies
- unusual increases
- duplicate-looking charges
- several potentially unused services
- a handful of pricing inconsistencies
That’s valuable.
But what does it prove?
It proves the AI found those opportunities.
It doesn’t necessarily prove that it systematically investigated every charge for every other type of savings opportunity.
- Did it determine whether every location was still active?
- Did it obtain utilization information that doesn’t appear on the invoices?
- Did it understand why each service was purchased?
- Did it compare configurations against actual requirements?
- Did it examine every applicable contract and amendment?
- Did it identify available discounts?
- Did it compare pricing against what similar customers are paying?
- Did it investigate alternative services?
- Did it know which information it didn’t have?
Unless we’ve explicitly designed for those questions, we may not know.
The danger isn’t necessarily that the AI gives you bad answers. It may give you several very good answers and still leave substantial money on the table.
AI has been trained to be helpful. Cost optimization sometimes requires something different.
There is another interesting dimension to this.
A lot of our interaction with AI is essentially summarization.
We give AI a large amount of information and ask:
“What matters?”
“What should I know?”
“What are the biggest issues?”
“What should I do?”
And we generally don’t want 173 answers.
We want the five things that deserve our attention.
That’s a feature.
But I’m not convinced that’s always the behavior we want when we’re trying to eliminate unnecessary expenditure.
If there are 47 legitimate opportunities to save money, I don’t necessarily want AI to decide which five are most interesting.
I want all 47.
I don’t want the system to think:
“I shouldn’t overwhelm the user.”
I want it to think:
“I have 47 opportunities. I will capture all 47, prioritize them appropriately, automate what can be automated, and involve a human only where human attention is required.”
That’s fundamentally different from summarization.
It’s systematic coverage.
You can’t analyze what you don’t know exists
The first requirement for Savings Coverage is almost embarrassingly basic.
Do we know everything we’re paying for?
Not just the vendors.
Not just the invoices.
Every service.
Every account.
Every line.
Every license.
Every location.
Every individual charge component.
That sounds obvious until you work with real enterprise data.
Bundled charges may hide individual services.
Legacy accounts can live outside the primary inventory.
Disconnected services may remain in billing systems.
New services get added without entering an established review process.
Acquisitions create disconnected datasets.
And the inventory itself may not contain the attributes necessary to evaluate whether something should exist.
If the system doesn’t know a service exists, it can’t optimize it.
Savings Coverage therefore starts with Inventory Coverage.
Having the charge isn’t enough. You need the context.
Suppose the inventory is complete.
Now we have another problem.
The invoice says we’re paying $2,000 a month for a service.
Is that good or bad?
We don’t know.
Perhaps the invoice tells us what was billed, but usage sits in a vendor portal.
The location information sits somewhere else.
The original contract is in a document repository.
A subsequent amendment changed the pricing.
The person who owns the service is listed in another system.
The reason the service exists isn’t recorded anywhere.
And the market pricing information needed to determine whether $2,000 is competitive is external to the organization entirely.
AI can’t reason from information it doesn’t have.
Which leads to a principle I think matters well beyond cost optimization:
Before asking AI to make a better decision, make sure it has the information required to make the decision.
And a good system should know the difference between “no opportunity exists” and “I don’t have enough information to determine whether an opportunity exists.”
Those are not remotely the same conclusion.
Imagine a system saying:
“I cannot complete the usage test for these 312 services because the required utilization data is not currently available.”
That’s useful.
The missing information has become visible.
It can now become a task.
Perhaps the system can retrieve it automatically.
Perhaps an integration needs to be built.
Perhaps a human needs to investigate.
But the test doesn’t silently disappear.
That’s Savings Coverage.
Every charge should face every relevant test
This is where I think AI becomes genuinely exciting.
Historically, we knew many of the questions we should ask.
The problem was applying them consistently.
Human experts can investigate an invoice.
They can look at usage.
They can understand a contract.
They can question why a service exists.
They can compare alternatives.
They can talk to vendors.
They can negotiate.
What humans are much less suited to doing is applying every relevant test to thousands or millions of charges, month after month, without missing anything.
It’s extraordinarily labor intensive.
So historically we prioritized.
We sampled.
We focused on large expenses.
We investigated exceptions.
We performed periodic optimization exercises.
We concentrated expert attention where we thought it would create the most value.
That was rational because human attention was the constraint.
AI changes that constraint.
If data collection and analysis can be automated, why shouldn’t every charge be verified every month?
That doesn’t mean a human manually reviews every charge every month.
Quite the opposite.
Human attention should be scarce. Savings coverage shouldn’t be.
The system should continuously do the repetitive work and escalate the exceptions that require judgment, investigation, negotiation or action.
“We checked it last year” isn’t enough
Cost optimization also isn’t a one-time event.
A service that was correctly optimized six months ago may not be correctly optimized today.
- Usage changes.
- Employees leave.
- Locations close.
- Business requirements change.
- Prices change.
- Discounts become available.
- Vendors introduce new plans.
- Contracts renew.
- Technology changes.
- New alternatives enter the market.
This is particularly obvious in dynamic categories.
Take mobile.
Usage, lines, plans, features, devices, pricing and promotions change continually.
Mobile needs to be optimized continuously.
Cloud infrastructure creates a similar problem. Something appropriately sized for one workload can become wasteful as utilization changes.
If the system is already collecting the information and the analysis is automated, why wait for an annual optimization project to discover something that became unnecessary nine months ago?
The objective should increasingly become:
Continuous Savings Coverage.
Every charge monitored.
Every relevant test applied at the appropriate frequency.
Every change evaluated.
Every exception captured.
Human intervention triggered when necessary.
Internal data can only tell you so much
There’s another place Savings Coverage breaks down.
Your internal systems know what you know.
Suppose the AI checks an invoice against the contract.
The contract says $100.
The invoice says $100.
Pass.
But what if comparable customers are paying $75?
What if a discount exists that isn’t in your current agreement?
What if the vendor introduced a new plan that would meet the same requirement for $60?
What if your existing configuration is commercially obsolete?
Nothing is “wrong.”
The invoice is correct.
The contract is being followed.
And you’re still overspending.
This is why external market intelligence matters.
- Pricing benchmarks.
- Available discounts.
- Vendor-specific knowledge.
- Alternative services.
- Commercial structures.
- Negotiation experience.
- Market changes.
A perfectly accurate invoice can still be a bad deal.
An AI system constrained entirely to internal information may never know that.
Then there’s the business context AI may not see
Not every apparent saving should be implemented.
A service may show no usage but exist for disaster recovery.
Two services may look duplicative but support different business requirements.
A more expensive configuration may exist because of a contractual, regulatory or operational requirement.
Eliminating one expense might create a larger expense somewhere else.
The person who owns the service may know something the data doesn’t.
That’s where human judgment remains essential.
But I don’t see this as humans versus AI.
I see human judgment as part of the system.
AI can perform the exhaustive analysis.
It can identify anomalies.
It can gather information.
It can quantify opportunities.
It can highlight where information is missing.
It can prioritize.
Then human expertise can be inserted precisely where judgment is valuable.
Human judgment becomes a control around AI, rather than a substitute for it.
That is a much more scalable model.
Finding the saving is only half the coverage problem
Let’s say the system works beautifully.
It identifies an unnecessary service costing $1,000 per month.
Great.
Have we saved $12,000?
No.
We’ve identified an opportunity to save $12,000.
Those are different things.
Now somebody may need to validate it.
The business owner may need to approve it.
The vendor may need to be contacted.
A disconnect order may need to be submitted.
The vendor may challenge the request.
The change needs to be scheduled.
The vendor needs to actually implement it.
Then the next invoice arrives.
Is the charge gone?
Perhaps it isn’t.
Perhaps the vendor agreed to disconnect it but didn’t.
Perhaps it was disconnected late.
Perhaps the wrong service was changed.
Perhaps a credit was promised but never appeared.
Perhaps a different charge was introduced.
This is what I think of as Realization Coverage.
The full process might look something like:
Opportunity identified → quantified → validated → stakeholder reviewed → approved → vendor contacted → change negotiated/requested → vendor implements → bill checked → saving confirmed → saving documented.
The opportunity isn’t finished when AI identifies it.
It’s finished when the financial result is verified.
This matters because an analytically perfect recommendation that nobody implements produces exactly the same realized saving as never finding it:
$0.
Implementation drift is real
Anyone who has spent enough time managing vendors knows that “the vendor agreed to do it” and “it happened correctly” are not the same thing.
That’s why accountability needs to be part of the system.
- Did the vendor complete the work?
- When?
- Was it completed correctly?
- Did the pricing change?
- Was the service disconnected?
- Did the expected credit arrive?
- Did the saving actually appear on the invoice?
- If not, what happens next?
Someone — human or automated — needs to keep pursuing the issue until the loop closes.
Otherwise the organization ends up with a growing list of theoretically valuable recommendations.
Which leads to a fairly simple test for AI cost-saving solutions:
Does the system reduce work required to realize savings — or simply generate more tasks for already overloaded people?
If it only produces recommendations, that’s useful.
But we should recognize what it is.
It may be an excellent opportunity-identification system.
It isn’t yet a complete savings system.
This changes how I think we should design AI for cost optimization
The obvious approach is:
Give AI our spending data and ask it to find savings.
I think the better approach is:
First understand all the ways savings are actually found and realized. Then design an AI + human system capable of achieving coverage across that entire process.
- Map the savings tests.
- Map the data required for each test.
- Identify where that data lives.
- Automate collection wherever possible.
- Continuously identify missing information.
- Apply every relevant test to every applicable charge.
- Bring in external market intelligence.
- Use humans where business context or judgment is required.
- Track every validated opportunity through execution.
- Verify the result against the actual bill.
Then repeat the above.
Every month.
That is a very different design philosophy.
And it starts with understanding the work before automating it.
The opportunity AI creates
None of this is an argument that AI isn’t ready for cost optimization.
I believe almost the opposite.
AI may finally allow us to address one of the fundamental constraints that has existed in this field for decades.
Coverage.
We have long known a huge number of questions that should be asked.
Historically, asking all of them, against everything, all the time, was extraordinarily labor intensive.
So organizations made trade-offs.
They focused expert attention where it was most valuable.
They reviewed periodically.
They investigated exceptions.
They accepted that perfect coverage wasn’t economically realistic.
AI and automation change that constraint.
We can increasingly build systems where the machine performs the exhaustive work and humans perform the valuable work.
The machine doesn’t get tired of checking the 48,731st charge.
It doesn’t decide a $14 discrepancy is too boring to investigate.
It doesn’t forget that a vendor promised a credit three months ago.
It can continuously look for changes in usage, pricing, inventory, contracts and market information.
And it can bring a human into the loop when judgment, negotiation, context or persistence is required.
That’s a much more interesting future than simply asking a chatbot to look at an invoice.
The standard should be higher than “AI found some savings”
When evaluating an AI cost optimization system, I’d ask questions like:
- Does it know every charge we’re paying?
- Does it know when information is missing?
- Can it automatically retrieve the billing, contract, portal, usage and inventory data it needs?
- Does it understand why services exist and who owns them?
- Does every charge face every relevant savings test?
- Does it continuously reevaluate those charges as circumstances change?
- Can it incorporate new market pricing and discounts?
- Does it distinguish an interesting anomaly from an actionable saving?
- Does it involve humans when business context or judgment is required?
- Does it help execute the opportunity?
- Does it hold vendors accountable for agreed changes?
- And does it return to the bill to prove that the saving actually occurred?
Those questions are much harder than:
“Can your AI find savings?”
But I think they’re the questions that matter.
Because the promise of AI isn’t that it knows some magical new way to save money.
It’s that we can finally apply everything we already know about finding savings to everything we spend, continuously.
That’s the opportunity.
Not a few clever recommendations.
Not a better summary of an invoice.
Not another dashboard full of opportunities for someone else to pursue.
A rigorous system in which every dollar can be interrogated, every relevant savings opportunity can be surfaced, human expertise can be applied where it adds value, and every valid opportunity can be followed until the saving is actually proven.
That’s what I mean by Savings Coverage.
And as more organizations decide to “throw AI” at cost optimization, I think it’s the standard we should be aiming for.
There is also a practical question worth asking:
How much Savings Coverage do you actually have today?
Can you account for every charge?
Do you have the data required to evaluate each one?
Is every charge being subjected to every relevant savings test?
Are you continuously incorporating changes in usage, pricing, discounts and market information?
And when an opportunity is found, can you prove it was followed through until the saving appeared on the bill?
We’ve developed a more detailed assessment process to help organizations answer those questions and identify where their current AI + human system has gaps in Savings Coverage.
And if you’d simply like to get a quick sense of where you stand, we’ve also created a basic self-assessment you can take here:
https://berlinpacific.com/quiz/
Because before asking “How much has our AI found?”, there may be a more important question:
“How much could we still be missing?”