Today’s mad dash toward AI continues to be defined by usage, but does usage mean value? We’ve seen enough research by now to know the answer. In one study, 80% of companies said they can’t tell if what they’re doing with AI is even working.
AI gives us an opportunity to re-examine long-held beliefs and models, but that doesn’t mean it automatically wins. There are reasons tried and true automation technologies haven’t gone anywhere. Years navigating this in the world of IT expense management (ITEM) – where the wrong call costs in more ways than one – have given Tangoe a unique lens and a proven framework for where AI earns its keep.
Before you hand over another process to a large language model, ask yourself, “Does AI truly belong here?” Here’s how Tangoe thinks of different AI and automation technologies: the criteria we weigh, the questions we ask, and the differences that drive value.
AI isn’t a monolithic “thing”
It’s a toolbox, and the payoff comes from knowing which tool fits which job. Tangoe breaks this down into four categories, keeping a fluid perspective because the landscape is quite fluid. What is true today will likely need a fresh look in six months, and so on.
- Automation/RPA (task repetition)
Best for: tasks that need to be done the same way, every time, with zero room for error – like pulling data off a website or processing an invoice. Tangoe’s automation processes invoices in 8 seconds, whereas it takes 5-10 minutes by hand. IBM does a deeper dive into RPA here.
At a glance:
- It rarely breaks on its own. When it does break, it’s usually because something external changed (like a website updated or a login screen moved).
- Explainability is built in because automation software follows exact instructions. It’s not necessarily important how the data was obtained, so long as it’s accurate.
- Machine Learning (pattern-spotting and predicting)
Best for: finding patterns in massive amounts of data, making predictions, and automating repetitive tasks that change over time. Tangoe’s ML crunches mountains of data to catch invoice errors being billed for, putting money back in the budget. This MIT article covers more on ML.
At a glance:
- It needs lots of repeated examples to learn from. If something breaks in a totally new, one-off way, ML can’t learn from that because there’s no past data to pull from.
- Where it helps most is learning human patterns. Here’s an example from the world of ITEM. A company’s automation is set up to check a vendor’s system daily for a new invoice. Most vendors post on schedule, but one runs a few days late here and there. Left on its own, the automation will keep checking and burn resources. Machine learning solves this by learning that vendor’s actual rhythm and then “telling” the automation exactly when to check instead of guessing. This is what we mean by a toolbox: AI and automation layered, working with one another and not in competition.
- Generative AI (the explainer)
Best for: taking complex information and turning it into something a person can understand and act on. At Tangoe today, it’s used for research (both shallow and deep), coding with humans-on-the-loop (HOTL), untangling large, unstructured data sets, writing and formatting, and smart, 24×7 assistance. Here’s a good explainer of GenAI from IBM.
FYI – Spectrum of self-determination of the AI:
- Human-in-the-loop (HITL) – does work but a person must approve
- Human-on-the-loop (HOTL) – does work but a person actively monitors
- Human-above-the-loop (HATL) – does work and has clearly defined boundaries and pre-determined checks, but no person actively monitors
For AI systems driving consequential outcomes, most organizations in 2026 continue to rely on HITL or HOTL approaches, leaving HATL largely theoretical in practice.
At a glance:
- When it comes to core data processing work that needs precision, it may or may not be suitable. GenAI is largely about interpretation: taking lots of information and turning it into something a person can understand easily.
- When the data is precise and carries a high level of risk, like in the world of ITEM, you intentionally don’t want interpretation and so GenAI use cases need to be chosen carefully.
- It’s great for, “Here’s what broke, here’s why, here’s what we tried, and here’s what to do next” scenarios – translating technical findings into plain language.
- Agentic AI (bots that act on their own)
Best for: taking complex, multi-step processes off your plate entirely – start to finish, without someone manually guiding each step. Learn more with this explainer from MIT.
At a glance:
- Trust comes down to accountability and explainability. Teams are afraid of the gavel falling on them for a decision that AI made. Managers don’t want the risk of defending a decision to a client they can’t explain.
- There are solutions built for this level of explainability and transparency, but buyer beware: according to Gartner, of the thousands of software vendors claiming to offer agentic AI, only about 130 are considered genuinely agentic. The vast majority are rebranded chatbots, RPA, or basic workflows wearing an “agentic” label – a practice commonly known as “agent washing.” It’s not surprising that Gartner predicts over 40% of agentic AI projects will be canceled by the end of next year. This is something Tangoe is very aware of. In our view, evaluating and trialing solutions – making sure they’re genuinely fit for purpose – takes just as much effort as execution. We work hard to make sure every investment is intentional, starting long before implementation.
- There’s a lot to consider in terms of costs. The more tokens agentic systems use, the more expensive they get. Gartner warns that agentic AI costs could balloon fivefold by 2028 – a serious problem, considering that 93% of companies are already exceeding their AI budgets.
- Also worth noting is a cost that most companies don’t budget for. In a recent McKinsey report titled, “Is that AI agent worth it?” the firm found that most of today’s agentic AI spend goes to refinement and correction cycles. In other words, dollars being burned by agents repeatedly doing quality control on their own work before committing to a final action.
We’re not here to discourage the use of agentic AI but simply point out the costs, risks, and trust gaps that exist.
How do you decide? Run your problem through 5 decision factors.
The goal is to apply the right tool, or the right combination of tools, to the problem. Start by running each problem through five decision factors.
- Cost. There are a few things to keep in mind here.
- Capability > price tag. Is this tool doing something to solve the problem that another tool genuinely can’t? Can it work at a scale human effort never could? If the answer is yes, the cost is most likely worth it. If the answer is no, you can probably do the same job cheaper with what you already have.
- AI costs move differently. Every time a new model ships, the previous generation typically gets cheaper. That means a tool that looks expensive today may not stay that way. Just like the landscape is fluid, so are the costs.
- Cost has to be measured against ROI continuously. If a project starts costing three times what it used to but only delivers 1.5 times the improvement, that’s a decision point: can you refine the approach to close the gap, or is it time to walk away? This is where the sunk cost fallacy does real damage. Teams have invested time, resources, and hope into a project, and that makes it hard to pull the plug – even when the numbers say you should. Knowing when to walk away is as important a skill as knowing when to invest in the first place.
Bottom line: The question isn’t “is this tool worth it today,” but whether you’re tracking capability, price, and ROI closely enough to know if that answer changes tomorrow, and whether you have the discipline to act on it if it does.
Learn step-by-step how to stop AI from blowing your budget with proven ROI practices →
- Reliability. Does the tool get the right answer every time and deliver it to the right place? This might not be most important for the problem at hand, but when it is, it’s table stakes. The moment a system starts delivering wrong answers, costs spike and productivity tanks.
Bottom line: It’s not just about uptime. It’s about accuracy and consistency.
- Scale. Do you need to run the same task repeatedly, at real volume, over time? Remember that if a tool charges per transaction, scale becomes a cost multiplier. A tool that looks cheap at 100 transactions can get expensive fast at 100,000.
Bottom line: Small-scale testing hides problems that only show up under real load (speed, cost-per-transaction, error rates compounding, etc.).
- Real-world resilience. Do you need to handle inconsistent data or unpredictable behavior from an outside party? If yes, real-world resilience matters. The more your process depends on something else acting consistently, the more critical real-world resilience becomes.
Bottom line: Think of certain AI tools as a resilience layer. Consider ML and RPA. If something goes wrong with the latter, the former can notice the pattern and switch to a pre-arranged alternative or fall back to a more manual process so data keeps flowing without pulling in an engineer.
- Long-term viability. A tool might be the cool new “thing” to use, and it might solve your problem today, but is it a long-term bet you’re willing to make? Can you replace it if you need to, and is there a pool of people who know how to maintain it? This is where it’s critical to keep in mind the fluidity of the landscape.
The bottom line: The newest, most capable tool might solve your problem well today, but you’re making a bet on something that could look completely different or not even be here a couple years from now.
Run real problems through these five factors enough times and the right tool tends to reveal itself.
Here’s what that looks like within Tangoe’s platform.
- Invoice processing needs reliable, rules-based accuracy – a.k.a. RPA. Machine learning doesn’t work because you don’t need a confidence score or a probability that something is right. You need 100% accuracy, every time. That also makes a large language model’s flexibility a liability because there’s zero room for interpretation. Reliability is the No. 1 priority here.
- Savings discovery needs pattern recognition at scale – a.k.a. ML. Spotting savings opportunities means finding patterns across huge volumes of spend and usage data that a person or a fixed rule would take too long to catch. Machine learning does this reliably at scale, and at a lower cost than more advanced AI tools. GenAI is also a great add-on here, taking those findings and translating them into plain language (what was found, why it matters, and what to do about it). That’s exactly why we built an AI assistant into our technology expense management (TEM) platform – to give customers instant financial insights in a way they can quickly understand and act on.
- Reporting and analytics need to be easy to understand – a.k.a. GenAI. The cost per token is worth it when you consider the time you’re saving outside of endless spreadsheets. You get the answers you need, explained in a way that allows you to be proactive.
- If we were to use agentic AI, it might look like semi-autonomous exception handling led by AI agents, with humans in the loop. For us, this is where real-world resilience and long-term viability still hold the line. Every invoice, order, and device carries real consequence, and the cost/trust math hasn’t tipped in favor of full autonomy yet.
Our final thoughts
You’re watching AI reshape the world and trying to figure out what it means for your organization. At the end of the day, it’s not about replacing everything with AI. It’s about adding the right capability where it makes sense.
When you run each problem through cost, reliability, scale, resilience, and viability, you’ll never end up with a guess. You’ll know you’ve got the right tool for the right job that will drive growth and value.
If you have questions, we’re here. If you want the full picture, get a closer look at how we apply AI and automation across our TEM platform, Tangoe One.