TRANSCRIPCIÓN 🔗
WEBVTT
1
00:00:00.140 --> 00:00:10.600
Sit through a procurement software demo this fall, and you'll see the same two words plastered everywhere: MCP-ready. I've watched rooms full of procurement leaders nod at that slide over the past few months.
2
00:00:11.100 --> 00:00:20.900
However, few of them could tell you what an MCP, or Model Context Protocol, actually means and what new business capabilities it enables for them and their teams.
3
00:00:21.680 --> 00:00:30.280
In short, it's a way to connect a large language model like Claude, ChatGPT, Mistral, or whatever else you're using directly to your procurement systems.
4
00:00:30.660 --> 00:00:36.420
It's incredibly useful to answer ad hoc queries, and once you go MCP, you don't come back.
5
00:00:37.040 --> 00:00:44.660
So today, I'm unpacking this concept to help you understand how this works, how to test vendors on their maturity, and what to do with it. Let's get into it.
6
00:00:45.940 --> 00:00:56.900
Technology presumes there's just one right way to do things, and there never is. AI will do what you tell it to do, but that may be very different from what you had in mind. Computers are useless.
7
00:00:57.000 --> 00:01:04.180
They can only give you answers. For a list of all the ways technology has failed to improve our quality of life, please press three.
8
00:01:06.860 --> 00:01:20.580
Welcome to ProcureTech Unpacked, the show for procurement professionals who want to master the technology shaping our function. I'm Joel Conneen Demers. Procurement already has plenty of acronyms: S2P, P2P, CLM, SRM.
9
00:01:21.160 --> 00:01:33.360
Unfortunately, the AI world just handed us one more, MCP, or Model Context Protocol. It shows up in LinkedIn posts, analyst updates, and vendor release notes, but usually without much explanation.
10
00:01:33.720 --> 00:01:37.320
Is it a product, a module, or something else that you have to license?
11
00:01:38.080 --> 00:01:51.440
Well, it's actually an open interface standard, and it decides how the AI tools that your company already pays for reach into your ProcureTech systems to grab supplier, contract, or PO data, so it's worth learning about.
12
00:01:52.460 --> 00:02:04.860
So today, four things: what an MCP is, where to draw the line between reading and writing, what MCP servers don't fix, and seven questions to ask vendors that claim that their solution is MCP-ready.
13
00:02:06.620 --> 00:02:13.160
Part one: What is an MCP? So as I mentioned, MCP stands for Model Context Protocol.
14
00:02:13.900 --> 00:02:30.940
Anthropic introduced it to the world in November twenty twenty-four, and it's an open interface standard that lets AI applications find and use external systems and data through one consistent interface instead of needing a custom integration for each system or even each API.
15
00:02:31.800 --> 00:02:42.200
Because it's an open standard, that means that it's a connection that is built once and works with any large language model that supports it. So let's think of this like a restaurant.
16
00:02:42.420 --> 00:02:56.900
Picture that your ERP or source-to-pay suite is a kitchen. Every dish that the kitchen can make for the outside customer is an API, like get supplier record, create a requisition, list open POs. Those are the meals.
17
00:02:57.920 --> 00:03:02.280
Now picture a diner who has never been to this restaurant. That is your LLM.
18
00:03:02.980 --> 00:03:11.800
With no menu, someone has to know what the kitchen makes, how to phrase each order, and what the kitchen will ask back, like, "How do you want that steak done?"
19
00:03:12.860 --> 00:03:25.380
Someone in your company has to hard code all that, creating something like a waiter, and when the kitchen changes a recipe, you need to update the waiter. MCP servers are a menu that directly connect to the kitchen.
20
00:03:26.040 --> 00:03:35.340
They list what's available, describe each dish in plain language, and group dishes so that the relationships between them are clear, such as appetizers, mains, and desserts.
21
00:03:36.220 --> 00:03:44.420
An AI agent reads the menu, picks what fits your question, and then places the orders in sequence to be able to answer you correctly.
22
00:03:45.320 --> 00:03:52.020
The kitchen still cooks, but the direct menu means that you no longer need a waiter who memorized everything to get you what you want.
23
00:03:52.860 --> 00:04:00.220
In procurement, that waiter is usually an expensive consultant who creates a custom interface between systems. So why does this matter?
24
00:04:00.920 --> 00:04:06.280
Well, before MCP, every AI tool had to be wired to every system and every data object separately.
25
00:04:07.160 --> 00:04:15.740
Say your company runs Microsoft Copilot, and you want it connected to four systems, your ERP, your contract system, supplier risk tool, and a sourcing tool.
26
00:04:16.540 --> 00:04:26.120
Well, for each one, you would map every object that you want from it, like suppliers, contracts, risk scores, or RFX events, and then you would build each integration.
27
00:04:26.960 --> 00:04:35.840
That's dozens of one-to-one integrations, and if you have multiple data objects that are the same across systems, you need to pull all of those and consolidate them together, right?
28
00:04:37.060 --> 00:04:42.700
And then each one breaks when the vendor changes its API, and you need to make updates to those interfaces.
29
00:04:44.400 --> 00:04:51.440
With MCP, if each one of those systems has an MCP server, that's four connections instead of the dozens that I just listed out.
30
00:04:52.040 --> 00:05:01.480
The LLM gets whatever each server exposes within your permissions and always uses the most up-to-date API in the background. So what does this enable?
31
00:05:02.320 --> 00:05:13.060
Imagine a category manager typing a q-question into your company Copilot. Which supplier contracts expire in the next ninety days and don't have a renewal owner, or at least
Load full transcript
not an active one?
32
00:05:13.820 --> 00:05:24.840
The LLM checks the connected MCP servers and reads the menu of available data objects. It finds a contract lookup in your source-to-pay suite. It picks that tool and fills the data range itself.
33
00:05:25.220 --> 00:05:34.360
That's the mapping your consultant or IT team used to do by hand. Then the server turns the request into an API call, and the suite sends back the results.
34
00:05:34.960 --> 00:05:42.960
The LLM looks up the owner on each contract and writes you the answer, and if contract owners are missing in the data, it'll tell you that too.
35
00:05:43.920 --> 00:05:50.300
So in this example, the category manager never named a system, didn't even know it existed, or doesn't have to know it existed.
36
00:05:50.660 --> 00:05:59.980
The MCP server told the LLM what data objects existed, the relationships between them, and the instructions for getting them out of the back-end source-to-pay suite via API.
37
00:06:00.996 --> 00:06:11.136
I use an MCP server for exactly this type of scenario, the ad hoc questions that used to mean exporting from five systems into Excel and then massaging the data during an afternoon.
38
00:06:12.436 --> 00:06:19.576
Once you have access to an MCP server, connecting actually takes about ten minutes. I couldn't believe how fast I got everything up and running.
39
00:06:19.576 --> 00:06:29.736
[whooshing sound] Part two: Where to draw the line between reading and writing. So an MCP server is very interesting, but it's only useful if it respects your user permissions.
40
00:06:30.296 --> 00:06:43.216
Your source-to-pay suite already knows what you can and can't see, what you can and can't do. Years of role design sit behind those rules, and an MCP server should only let you act as you with your accesses.
41
00:06:43.756 --> 00:06:56.236
Otherwise, a coordinator checking a contract expiry gets the CFO's view of supplier pricing or the opposite use case, the MCP server sees less than you do and answers come back incomplete and people stop trusting the tool.
42
00:06:56.956 --> 00:07:07.256
So you need three things for an MCP server to be useful. First, your role sets the limit. Through the MCP server, you can only do what you can do manually in the backend system.
43
00:07:07.696 --> 00:07:20.456
Nothing more, nothing less, and every action traces back to a named user. Second, you decide what you hand over. When you connect an MCP server via LLM, you pick which meals the LLM gets access to.
44
00:07:20.556 --> 00:07:32.736
So maybe it reads your contracts, but it never gets to change them. Third, on that same note, you draw the line between reading and writing. So as you're launching, maybe MCP server is read-only, right?
45
00:07:32.776 --> 00:07:38.936
It's just getting data and helping you do ad hoc analysis. However, it doesn't allow you to come back and make changes in the system.
46
00:07:40.096 --> 00:07:51.116
You can then migrate to using the LLM to
do write actions, right, with approval and, uh, making sure that there's a human in the loop as you watch the thing behave for the first few weeks and months.
47
00:07:52.156 --> 00:08:04.556
That same rule will cover AI agents because as companies set up agents as full users with their own logins, their own roles, their own permissions, those agents can reach the data through the same MCP server.
48
00:08:05.156 --> 00:08:09.896
So a sourcing agent with buyer permission, an invoicing agent with read access to invoices.
49
00:08:10.516 --> 00:08:16.996
You govern them like you would any other user by provisioning them, reviewing their accesses, and switching them off when they're no longer needed.
50
00:08:17.516 --> 00:08:26.096
But this only works if you've got really sophisticated permission management and permission setup in the backend system to avoid undesirable behaviors.
51
00:08:26.596 --> 00:08:39.876
Having used MCP servers for a while now myself, this is a really cool way to reduce the time you spend on ad hoc analysis, and it also enables you to answer a whole bunch of new questions that you had never dared to answer because of the work involved.
52
00:08:39.876 --> 00:08:51.196
[whooshing sound] However, part three: What MCP servers won't fix. MCP servers connect an LLM or an AI agent to your data, but it does nothing for the quality of that data.
53
00:08:51.936 --> 00:09:04.956
So if you ask an LLM for the total spend for a global supplier that lives under five different vendor IDs in five different systems, it might return five partial answers or make some mistakes in the consolidation process.
54
00:09:06.036 --> 00:09:15.756
Yes, the menu is standardized, but the food remains whatever your master data team cooked up in the backend system. That makes MCP a good place to start on data quality, actually.
55
00:09:16.456 --> 00:09:26.656
Remember the missing contract owners in my earlier example? That MCP request can show you that gap pretty fast. And MCP servers won't replace your structured integrations either.
56
00:09:27.156 --> 00:09:37.156
MCP shines when the questions change every time, like an ad hoc spend question, a contract lookup, an analysis across systems. But a PO exchange, for example, is the opposite.
57
00:09:37.316 --> 00:09:44.216
It's the same document, the same fields over and over again, and everyone has to land exactly right with the supplier to avoid issues.
58
00:09:44.856 --> 00:09:56.776
That job still belongs to EDI, CXML, supplier network connections, traditional structured integration. You don't want an LLM reading a menu to decide how to get a purchase order to a supplier.
59
00:09:57.996 --> 00:10:04.316
MCP also doesn't decide when a tool gets used. The LLM model and the application around it make that call.
60
00:10:04.776 --> 00:10:15.916
So you still need testing, guardrails, clear rules on which decisions stay human and critical process steps that need to be respected. Furthermore, someone has
to own each server.
61
00:10:16.336 --> 00:10:27.156
When your CLM vendor changes an API, they need to update their server. Otherwise, the agent starts ordering dishes that no longer exist because they're still on the menu, but the kitchen doesn't cook them any longer.
62
00:10:28.136 --> 00:10:35.936
Then there's security. MCP gives an AI agent or LLM a path into the systems that hold your supplier contract, pricing, and bank details.
63
00:10:36.516 --> 00:10:48.236
If an agent can read a supplier's PDF and also has write tools, then text hidden in that file from the supplier can steer an agent to do th-- wrong things or nefarious things in your systems.
64
00:10:48.936 --> 00:10:55.496
So this is a new cybersecurity threat at the same time. A menu lists the dishes, but it doesn't check for your allergies.
65
00:10:55.996 --> 00:11:04.116
Permissions and role design as well as system and process architecture still remain very important if you're gonna chase the capabilities that this tool gives you.
66
00:11:04.136 --> 00:11:16.056
[whooshing sound] Part four: Seven questions for the vendors. So back to that initial slide that everybody's been seeing. MCP-ready can mean a production server. It can also mean a slide.
67
00:11:16.516 --> 00:11:28.336
So here's seven questions to test maturity as you get educated on MCP servers with your software vendors. Number one, is your MCP server generally available today or is it a demo or roadmap item?
68
00:11:29.036 --> 00:11:39.936
Number two, which actions does it expose? Which of them can write, change, or approve in the backend system? Number three, does it inherit each user's existing permissions? Is it open?
69
00:11:40.076 --> 00:11:51.936
Does it support individual AI agent permissions as well or only the humans? Four, can we switch off individual actions on a user-specific basis or not?
70
00:11:52.636 --> 00:11:59.676
Number five, what g- gets logged and can we trace an action back to a specific user and AI agent? What does that look like?
71
00:12:00.636 --> 00:12:13.416
Number six, who hosts and maintains the server and what happens when the APIs change un- beneath the server? Number seven, which LLMs can connect to it and can we use the one IT already approved in-house?
72
00:12:14.456 --> 00:12:21.376
And then if the answers to three and five are vague, you've learned what you needed to know. They are not, in fact, MCP-ready.
73
00:12:22.256 --> 00:12:29.056
MCP servers are just a menu that let LLMs and AI agents find and use the systems and data that you already own.
74
00:12:29.576 --> 00:12:43.656
It makes agents much easier to con-connect to your systems, and it also puts permissions, data quality, and ownership under a bright light because AI will use whatever it can reach, as we've seen in the news with the Hugging Face ex-- incident, for example.
75
00:12:44.896 --> 00:12:51.536
Start with read-only, ask the hard questions, and fix your data in parallel to enable the capabilities and insights you're looking for.
76
00:12:52.336 --> 00:13:01.856
The written version of this topic with the full vendor checklist is in the Pure Procurement Newsletter. I will leave a link to the edo- edition where we covered this in the show notes. See you next time.
77
00:13:01.856 --> 00:13:04.776
[upbeat music] That's a wrap on this episode of ProcureTech Unpacked.
78
00:13:05.176 --> 00:13:14.996
If you enjoyed this one, give us a follow on your podcast app or YouTube and sign up for the Pure Procurement Newsletter to get procurement technology insights delivered straight to your inbox.
79
00:13:15.576 --> 00:13:25.456
Otherwise, what's the single best way to tell us you're enjoying the show? Leave us a five-star review on Apple Podcasts or Spotify, like the video on YouTube, or drop us a comment.
80
00:13:25.956 --> 00:13:36.616
Whatever platform you're listening on, your feedback means absolutely everything to a niche show like this. All the links, reports, and resources we mention in the episodes are in the show notes.
81
00:13:37.076 --> 00:13:52.616
And as always, a big thank you to our production team over at Resonate Recordings for making us look and sound good. I'm Joël Conneau de Mès. We'll see you next week. [outro jingle] Resonate.